NeuroFly
Evidence · release 2.0.0 · runs of 23 September 2026

Validation and verification

Two different questions. Validation: does the model reproduce what is known about real flies — and for the right, circuit-level reasons? Verification: does the software do exactly what we say it does? Both are reported here in full, including where the model fails.

How the experiments were run

Each benchmark is one of the ten guided experiments built into NeuroFly, run in a bare arena with no instruments attached, at a fixed time of day. Every trial starts from a freshly reset network and body with its own seed drawn from one master seed (20260923), so the complete set can be re-run and reproduced exactly. Proportions are shown with Wilson 95% confidence intervals; group differences are tested with Fisher’s exact test or the Mann–Whitney U test.

BenchmarkFinding in real fliesModelResult
Looming escapeLooming drives the giant fiber; escape probability rises with looming strengththreshold response via the giant fiberReproduced
Looming detectorsLC4 and LPLC2 carry complementary looming features to the giant fiberboth needed; LC4 dominantReproduced
Wind vs soundseparate Johnston’s-organ populations; wind does not trigger escapewind: no escape; sound: escape (prediction)Reproduced
Taste decisionsugar drives proboscis extension; bitter suppresses itgraded extension, full bitter vetoReproduced
Antennal groomingJO-F activation elicits antennal grooming; DNg12 drives head sweepsdust-dependent; needs JO-F and DNg12Reproduced
Inhibition gates escapegiant-fiber recruitment is shaped by excitation–inhibition balancedose–response: escape 100% → 0%, half-maximal at 1.87× inhibitionReproduced
Activation screenten cell types with published activation phenotypes6 of 10 phenotypesPartial
Habituationthe escape pathway habituates to repetitionno habituationNot reproduced
Associative learningmushroom-body, dopamine-gated learningno learningNot reproduced
Thermal preferenceflies gather near 24–25 °C in a gradientno preferenceNot reproduced

Benchmarks that reproduce

Hover over a data point for its value, interval and sample size; every figure also has its data table.

Figure 3 · Looming escape threshold

Both eyes see a looming stimulus for 300 ms at nine intensities; does she take off?
von Reyn et al. 2014; Card & Dickinson 2008

All-or-none threshold at intensity 0.14 (logistic fit). 0/12 takeoffs at 0.04, 12/12 from 0.2 upward; giant-fiber latency 5 ms at the strongest stimulus. Takeoff only ever followed a giant-fiber spike.

Data table
IntensityTakeoff95% CIn
0.040%0–24%12
0.080%0–24%12
0.120%0–24%12
0.1692%65–99%12
0.20 – 1.00100%76–100%12 each

Figure 4 · Which looming detectors drive escape?

Near-threshold loom with LC4, LPLC2 or both silenced (15 trials each).
von Reyn et al. 2017; Ache et al. 2019

Silencing LPLC2 reduces giant-fiber spikes by 42%, LC4 by 83%, both by 99.8% (each p < 0.001 vs intact, Mann–Whitney). Takeoff: 100%, 100%, 93%, 7%. Both populations contribute, as in real flies.

Data table
ConditionGF spikes (mean ± SEM)Takeoffn
intact35.9 ± 0.8100%15
LPLC2 silenced20.8 ± 0.2100%15
LC4 silenced5.9 ± 0.293%15
both silenced0.07 ± 0.077%15

Figure 5 · Wind and sound reach different neurons

Equal-strength wind, near-field sound or an air puff (both), 12 trials each.
Kamikouchi et al. 2009; Yorozu et al. 2009

Wind: 0 giant-fiber spikes, 0/12 takeoffs — as in real flies, which stop rather than flee in wind. Sound: 45 spikes, 12/12 takeoffs (Fisher p < 0.001), because only the auditory JO-A/B neurons contact the giant fiber in the connectome. That sound triggers escape is a model prediction, not an established finding.

Data table
StimulusGF spikes (mean ± SEM)Takeoffn
none00%12
sound45.0 ± 0.2100%12
wind00%12
air puff46.4 ± 0.2—12

Figure 6 · Sugar, bitter and the proboscis

1 s of sugar on the labellum at seven concentrations; then sugar 0.75 with rising bitter (8 trials each).
Shiu et al. 2024

(a) Sugar alone

(b) Bitter added to sugar 0.75

Proboscis extension needs sugar above ~0.68 (logistic 50% point). Bitter from 0.5 upward abolishes it: 88% → 0% (Fisher p = 0.0014) — the veto runs through the measured gustatory relay neurons, not through a rule.

Data table
ConditionExtension95% CIn
sugar 0 – 0.50%0–32%8 each
sugar 0.75100%68–100%8
sugar 1.088%53–98%8
sugar 0.75 + bitter 0 / 0.2 / 0.3588% / 100% / 88%8 each
sugar 0.75 + bitter 0.5 – 1.00%0–32%8 each

Figure 7 · Dust and head grooming

Six amounts of dust on the antennae for 2.5 s; then full dust with JO-F or DNg12 silenced (8 trials each).
Hampel et al. 2020; Guo, Zhang & Simpson 2022

Head grooming rises with dust (25% at 0.3, 100% from 0.45). Silencing either the JO-F mechanosensors or the DNg12 descending neurons abolishes it: 8/8 → 0/8 (Fisher p < 0.001). The grooming stops once the dust is gone.

Data table
ConditionHead grooming95% CIn
dust 00%0–32%8
dust 0.325%7–59%8
dust 0.45 – 1.0100%68–100%8 each
dust 1.0, JO-F silenced0%0–32%8
dust 1.0, DNg12 silenced0%0–32%8

Figure 8 · Inhibition as a dose–response

A near-threshold loom (0.15) while the strength of all inhibitory synapses (GABA and glutamate class) is scaled from 0.25× to 4× normal, 10 trials each.
von Reyn et al. 2014

Escape falls from 100% to 0% as inhibition is strengthened; the half-maximal inhibitory gain is 1.87× normal (logistic fit on log gain). Giant-fiber spikes fall from 12.3 at 0.25× to 0 at 4× (slope −3.2 spikes per unit gain, p < 0.001): feedforward inhibition decides whether the looming signal recruits the escape neuron.

A dose–response on transmitter classes — the kind of readout an in-silico pharmacology needs, not a model of any specific drug.

Data table
Inhibitory gainTakeoff95% CIn
0.25× – 1.5×100%72–100%10 each
2×10%2–40%10
3×10%2–40%10
4×0%0–28%10

Activation screen: 6 of 10 phenotypes

Each identified population is driven for 300 ms (6 trials) and the behaviour that appears beyond an unstimulated control is compared with the published activation phenotype.

ActivatedPublished phenotypeModel
Giant fiber (DNp01)escape takeoff Lima & Miesenböck 2005takeoff 6/6Match
LPLC2escape takeoff Ache et al. 2019takeoff 6/6Match
MDNbackward walking Bidaye et al. 2014backward 6/6Match
DNg12head grooming Guo et al. 2022head grooming 5/6Match
JO-Fantennal grooming Hampel et al. 2020head grooming 6/6Match
Sugar receptor neuronsproboscis extension Shiu et al. 2024extension 6/6Match
DNp09forward walking Bidaye et al. 2020no walking beyond controlMismatch
DNg11front-leg rubbing Guo et al. 20221/6Mismatch
DNa01/02, leftturning left Rayshubskiy et al. 2025no turning; backward walkingMismatch
DNa01/02, rightturning right Rayshubskiy et al. 2025no turning; takeoffMismatch

The mismatches concern walking and steering commands, whose effect in the model passes through the modelled brain–nerve-cord interface and body rules — the least constrained part of the model. They mark where the next model work has to go.

What the model does not reproduce

Negative results are part of the evidence. These three findings from real flies are not reproduced by NeuroFly 2.0.

Habituation

Real flies’ giant-fiber pathway habituates to repeated stimulation (Engel & Wu 1996). In 20 repetitions the model’s response did not decline; with the optional plasticity rule it increased (sensitisation; slope +0.45 spikes per repetition, p < 0.001). The synaptic depression that likely underlies habituation is not modelled.

Associative learning

Fly learning is dopamine-gated in the mushroom body (e.g. Ueno et al. 2017). The generic timing rule available in NeuroFly produced no learning (index 0.07 paired vs 0.04 reversed, p = 0.19). The mushroom body is not part of the running circuit.

Thermal preference

Real flies gather near 24–25 °C in a gradient (Sayeed & Benzer 1996). The model spent no more time in the comfort zone than in a flat arena (15% vs 16%, p = 1.0). Warmth avoidance depends on internal thermosensors (Hamada et al. 2008) that are not in the circuit.

Verification

28 automated test suites run the real model — no mocks — and must all pass before a release. Release 2.0.0: all pass. A further end-to-end test starts the complete application and checks every workspace.

Among the invariants they enforce:

  • The giant fiber is silent over 4 s of rest and fires within about 10 ms of an abrupt loom.
  • Wind produces essentially no giant-fiber spikes; sound of the same strength produces hundreds — from the wiring alone.
  • Temperature reaches only the thermoreceptor neurons, never the visual looming detectors.
  • The nerve cord’s ascending neurons measurably change brain activity when the locomotor circuit is driven — the feedback path is connected, not merely present.
  • The taste and grooming pathways respond in a graded way and fall silent after the stimulus.
  • Two runs with the same seed are identical; statistics match textbook values.
  • The display can never change what the fly’s own eye receives.

Data fingerprints

SHA-256 hashes of the data files shipped with release 2.0.0. NeuroFly recomputes them at start-up and writes them into every recording, so any result can be tied to the exact data it came from.

FileSHA-256
circuit.json2c25cae5671eedcd11133d4e7d0fc278
bb43bf29780c72cd0a119073975745d7
thermo_extension.json2cafb852f477d9f21433e8e3563ca9e3
089f44d62d43d36f2cefd4dcc630847d
sensory_extension.jsonef1125dbfaea25899dca21d7f1d59813
6ae1ad1ac4fe370d7845083eef730746
locomotor_circuit.jsone2060c02594f3a527c98092c5daad0c0
fce7c955460355e47d969b0d4f5c0ed1
circuit_annotations.json00967643eb9a0b5e908c10941949af98
27206c9da5a7c7978b1620636ad8ad24
brain_points.json0b26b67d1faa2d380c057c501408b632
7e03fb85c4c9dabc987a5933cc92f5da

Status of this evidence

These results have not yet been peer-reviewed. They are produced by the software itself, from published data, and can be re-run by anyone with the application. We welcome scrutiny: if you find an error — in the data handling, the statistics, the literature, or the interpretation — please tell us.

Report a scientific error

How to cite

NeuroFly (2026). An in-silico fruit fly on measured connectomes. Version 2.0.0. https://neurofly.app Please also cite the datasets it uses: FlyWire: Dorkenwald et al. 2024; Schlegel et al. 2024 (Nature 634). MaleCNS: Berg et al. 2025 (bioRxiv 2025.10.09.680999).

References

  1. Ache JM, et al. Neural basis for looming size and velocity encoding in the Drosophila giant fiber escape pathway. Curr Biol 29, 1073–1081 (2019). doi:10.1016/j.cub.2019.01.079
  2. Bidaye SS, et al. Neuronal control of Drosophila walking direction. Science 344, 97–101 (2014). doi:10.1126/science.1249964
  3. Bidaye SS, et al. Two brain pathways initiate distinct forward walking programs in Drosophila. Neuron 108, 469–485 (2020). doi:10.1016/j.neuron.2020.07.032
  4. Card G, Dickinson MH. Visually mediated motor planning in the escape response of Drosophila. Curr Biol 18, 1300–1307 (2008). doi:10.1016/j.cub.2008.07.094
  5. Engel JE, Wu C-F. Altered habituation of an identified escape circuit in Drosophila memory mutants. J Neurosci 16, 3486–3499 (1996). doi:10.1523/JNEUROSCI.16-10-03486.1996
  6. Guo L, Zhang N, Simpson JH. Descending neurons coordinate anterior grooming behavior in Drosophila. Curr Biol 32, 823–833 (2022). doi:10.1016/j.cub.2021.12.055
  7. Hamada FN, et al. An internal thermal sensor controlling temperature preference in Drosophila. Nature 454, 217–220 (2008). doi:10.1038/nature07001
  8. Hampel S, et al. Distinct subpopulations of mechanosensory chordotonal organ neurons elicit grooming of the fruit fly antennae. eLife 9, e59976 (2020). doi:10.7554/eLife.59976
  9. Kamikouchi A, et al. The neural basis of Drosophila gravity-sensing and hearing. Nature 458, 165–171 (2009). doi:10.1038/nature07810
  10. Lima SQ, Miesenböck G. Remote control of behavior through genetically targeted photostimulation of neurons. Cell 121, 141–152 (2005). doi:10.1016/j.cell.2005.02.004
  11. Rayshubskiy A, et al. Neural circuit mechanisms for steering control in walking Drosophila. eLife 13, RP102230 (2025). doi:10.7554/eLife.102230
  12. Sayeed O, Benzer S. Behavioral genetics of thermosensation and hygrosensation in Drosophila. PNAS 93, 6079–6084 (1996). doi:10.1073/pnas.93.12.6079
  13. Shiu PK, et al. A Drosophila computational brain model reveals sensorimotor processing. Nature 634, 210–219 (2024). doi:10.1038/s41586-024-07763-9
  14. Ueno K, et al. Coincident postsynaptic activity gates presynaptic dopamine release to induce plasticity in Drosophila mushroom bodies. eLife 6, e21076 (2017). doi:10.7554/eLife.21076
  15. von Reyn CR, et al. A spike-timing mechanism for action selection. Nat Neurosci 17, 962–970 (2014). doi:10.1038/nn.3741
  16. von Reyn CR, et al. Feature integration drives probabilistic behavior in the Drosophila escape response. Neuron 94, 1190–1204 (2017). doi:10.1016/j.neuron.2017.05.036
  17. Yorozu S, et al. Distinct sensory representations of wind and near-field sound in the Drosophila brain. Nature 458, 201–205 (2009). doi:10.1038/nature07843