---
language:
- en
pretty_name: FlyWire-informed fly odor-reward simulation — V4b
tags:
- neuroscience
- drosophila
- flywire
- mujoco
- simulation
- associative-learning
configs:
- config_name: v4b_choices
  default: true
  data_files:
  - split: calibration
    path: tables/v4b_calibration_choices.parquet
  - split: validation
    path: tables/v4b_validation_choices.parquet
- config_name: v4_failed_calibration
  data_files:
  - split: calibration
    path: tables/v4_calibration_choices.parquet
---

# FlyWire-informed odor-reward simulation: individual-behavior V4b

**The full predeclared validation FAILED. This is synthetic simulation data,
not measured fly behavior or a quantitative reproduction of Kaun et al. (2011).**

[Detailed results](results/v4b/RESULTS.md) ·
[Code and protocols](https://github.com/histochemichael/fly-sud-simulation)

![Paper versus simulation](results/v4b/paper_vs_sim_v4b.png)

## Findings and limitations

256 independently seeded validation flies, four conditions (paired, unpaired,
untrained, retrieval-DAN-silenced), two delays (30 min, 24 h), 32 flies per cell:
8 reciprocal replicate pairs with 2 flies per odor group.
Paired CPI was -0.59375 ± 0.13311 SEM early and +0.46875 ± 0.15264 late.
Paired-versus-unpaired exact two-sided paired sign-flip tests gave Holm-adjusted
p=0.03125 and p=0.09375. The late contrast failed the declared statistical gate.
Ten stationary nonresponders remain in denominators. No selective exclusions
or post-validation tuning were performed.

Both timing points informed calibration: independent validation seeds assess
model repeatability, not held-out time-course prediction or biological validity.
Individual variation is modeled, not fitted to measured individual flies.
The memory transition remains too early; 7-day persistence and shock resistance
are not reproduced. The paper and simulation use different tests and group sizes.
CPI SEM is across reciprocal pairs, not a standard deviation of individual paths.

## Included runs

- V4 failed calibration: 48 trials at gains 5, 20, 80, using the same 16 seed
  profiles across gains (common random numbers). No candidate passed.
- V4b calibration: 32 new flies at predeclared gain 50.
- V4b validation: 256 new flies after freezing the selected configuration.

Failed calibration records are preserved. Earlier V1/V2/V3 experiments are not
pooled into these tables. Original V1 archives remain in the GitHub release.

## Data layout and reuse

Dataset Viewer exposes only explicit individual-choice tables with stable
individual_key = version/batch/fly_id. Full records are under raw/v4/ and raw/v4b/.
Each trial includes Parquet state/physics/learning histories, summary, individual
configuration, replay and NPZ naive/post-training/pre-test/final learning states,
including exact FlyWire root IDs. Configuration, seeds, software versions and
source snapshots are preserved alongside each run. MANIFEST.json provides
SHA-256 checksums for all uploaded files except the manifest itself.

~~~python
from datasets import load_dataset
choices = load_dataset("Histochemichael/fly-sud-simulation", "v4b_choices")
from huggingface_hub import snapshot_download
folder = snapshot_download("Histochemichael/fly-sud-simulation", repo_type="dataset")
~~~

Physics state indices join within the same trial, not globally. Nonresponse
latencies are censored at the 120-second assay limit, never zero-latency choices.
At an absorbing collection zone physics stops; subsequent hold rows are not
walking. Do not invent trajectories beyond collection. Test memory states can
be reconstructed from initial state, elapsed time and archived update equations.

## Model and compute

Real FlyWire FAFB v783 connectivity informs an 8,570-neuron olfactory/mushroom-body
subset. All simulated flies share that reference anatomy; individual connectomes
were not measured. Assumed dynamics, noise, memory kinetics and motor decoding
are not determined by anatomy. Dopamine is a simplified DAN-to-MBON retrieval
gate; this is not mechanistic DAN-to-KC plasticity or a whole-brain simulation.
Learning is online associative/reward-modulated, not deep learning or a complete
biological addiction model.

Testing uses actual MuJoCo/NeuroMechFly physical dynamics. Conditioning is
stationary. Local Intel Core i7-9700F, eight CPU workers; no GPU acceleration.
Publishing dependencies may differ from archived experiment environments.

## Sources and rights

Kaun et al. (2011), A Drosophila model for alcohol reward, Nature Neuroscience,
DOI:10.1038/nn.2805. [Primary source](https://pmc.ncbi.nlm.nih.gov/articles/PMC4249630/).
The user-supplied paper figure is retained unchanged for research comparison,
not a training example. See [attribution](results/v4b/PAPER_ATTRIBUTION.md).
Original author/publisher rights remain; no blanket license is asserted.
FlyWire source connectivity is attributed under its CC BY 4.0 source terms:
https://zenodo.org/records/10676866 . Branding and other third-party assets
retain their own rights. Do not infer an unrestricted license from public access.