FLYBID · METHODS
AboutReturn to the neuron board
Methods · protocol · provenance

Train on cVA. Validate on MN9.

The live homepage sells 165,122 reconstructed FlyEM neurons on a rotating connectome. This page is the assay record. The two are not the same claim.

CLAIM BOUNDARY
MB sugar conditioning itself is prior art. No biological-learning or courtship-learning claim until frozen-MB and unpaired-US controls fail to match the paired arm. No transfer-success claim until all 99 mapped MN9 types complete the frozen gate.

Primary experiment: male pheromone training

Template EXP-C-CVA. Pair DoOR odorant 11-cis vaccenyl acetate (cVA / Or67d / DA1) as CS+ against 3-octanol as CS−. One sugar dopamine pulse at each relevant CS+ block onset depresses eligible KC→MBON synapses. Frozen-MB keeps mushroom-body readout and disables plasticity. Unpaired-US delivers the same pulse on CS−.

Live neurons are measured FlyEM somaLocation points, colored by superclass and by mushroom body, fruitless / doublesex, ORN_DA1, and identified descending / MN9 motor neurons. Fruitless and DA1 spike counts are reported every telemetry tick. Occupancy preference is single-agent dwell time, not the Tully–Quinn group-choice index.

The increment is the male-specific odor and MaleCNS sex-circuit readout, not a new plasticity rule. Computational seeds are not biological replicates.

Supporting question: MN9 transfer

Supporting endpoint: cell-type-level agreement between external experimental rostrum-extension labels and MaleCNS MN9 responses. This checks whether the same unmodified model still predicts published SEZ phenotypes.

The contribution is a direct transfer test: preserve the Shiu et al. connectivity-only LIF logic, move it to the independently reconstructed MaleCNS v1.0 connectome, prohibit benchmark calibration gains, and quantify agreement against experimental behavior. If the prespecified gate passes, the same system can prioritize uncertainty-robust candidates for wet-lab testing. Candidate output remains a prediction, not evidence of phenotype.

Auditable data chain

01External outcome labels. Shiu et al. 2024 Nature Supplementary Table 3 supplies 106 named cell types and experimental rostrum-extension rates. A rate greater than zero is the positive label.
02Historical identities. Shiu’s v630 FlyWire root IDs are read from the published identity resource and converted once from pickle to inert JSON.
03Current FlyWire annotation. Only an identical root ID surviving into v783 is accepted. FlyWire v3.1.0 supplies its curated cell type.
04MaleCNS mapping. Curated FlyWire cell type is matched exactly to MaleCNS v1.0 flywireType. Fuzzy aliases and manual rescue are prohibited.
05Simulation endpoint. Every MaleCNS neuron carrying the mapped type is stimulated; left and right MN9 firing rates are recorded.
06External comparison. MaleCNS prediction is compared with the held-out experimental label to form the confusion matrix.

Identity mapping and exclusions

Coverage is 99 / 106 cell types. Seven rows are excluded by the deterministic identity policy, before simulation. Exclusion is not treated as a negative prediction.

viceFlyWire v3.1 cell type has no MaleCNS v1 flywireType match.
TH_VUMNo v630 root ID survives as an identical v783 root ID.
bambooNo v630 root ID survives as an identical v783 root ID.
gallinuleNo v630 root ID survives as an identical v783 root ID.
meteorNo v630 root ID survives as an identical v783 root ID.
Salivary_MN13No v630 root ID survives as an identical v783 root ID.
weaverFlyWire v3.1 cell type has no MaleCNS v1 flywireType match.

Fixed activation protocol

Drive50 Hz
Duration100 ms
Repetitions10
Contact efficacy0.275 mV

Each mapped cell type is run independently. All matching MaleCNS neurons receive 50 Hz drive for 100 ms. Ten deterministic-seed repetitions are averaged by MN9 side. The connectivity-only model uses a fixed 0.275 mV per synaptic contact. Passive LIF parameters and contact efficacy remain fixed across cell types.

No calibration gains are used for the benchmark (gains=None). Existing calibrated gains used elsewhere in the project are excluded from this transfer test. Partial output is checkpointed row by row, but partial scores do not open the claim gate.

Prediction and scoring

experimental_positive = experimental_rostrum_extension_rate > 0
male_cns_predicted_positive = max(mean MN9 left Hz, mean MN9 right Hz) > 0

The complete 99-row screen yields true positives, true negatives, false positives, and false negatives. Accuracy, sensitivity, and specificity are reported without hiding error rows. False positives and false negatives are biologically informative transfer failures, not display exceptions.

The behavior labels are external to MaleCNS and are not inferred from the simulation. The benchmark does not optimize a threshold, contact gain, mapping, or model parameter against those labels.

Claim gate

Completion is necessary, not sufficient

  1. All 99 pre-mapped cell types must finish under the frozen protocol.
  2. Mapping coverage, seven exclusions, full confusion matrix, and all error rows must remain visible.
  3. Transfer performance must be judged against the preregistered acceptance criterion; the interface does not invent or move that threshold.
  4. Only a passing complete screen permits conditional wet-lab candidate prioritization. It does not validate those new candidates; experiments must do that.

Anatomy and control assays

Measured anatomical projection

The observatory reads binary soma coordinates from /neurons: three float32 values per sampled neuron followed by one uint8 group value. X-Count defines row count. Groups are central brain, optic lobe, ventral nerve cord, mushroom body, fruitless / doublesex, DA1 pheromone ORNs, and identified motor / descending neurons. Every displayed point is a measured soma coordinate. No skeletons or invented silhouettes are shown.

Optomotor assay

A cylindrical square-wave grating has 12 cycles, contrast 0.85, rotates at ±30°/s, and reverses every 10 s. Optomotor index measures turning with versus against the drum, excluding absolute heading changes below 2°. DNa02 silencing zeroes outgoing synapses and restores them after the run.

Olfactory T-maze

The default live pair is cVA versus 3-octanol. The older OCT/MCH pair remains available as --pair octanol. A fixed dopamine pulse at each relevant CS+ block onset gates Kenyon-cell-to-MBON depression. This pulse is a computational proxy; sugar sensory nerves and dopamine dynamics are not simulated. Frozen-MB and unpaired-US runs are controls. The metric is single-agent occupancy preference, not the Tully–Quinn group-choice performance index.

These assays are the training experiment and its controls. The MN9 transfer screen cannot open a learning claim, and a training run cannot open the MN9 claim gate.

Limitations

  • MaleCNS and the female FlyWire-derived benchmark differ by sex, reconstruction, annotation release, and identity coverage. Transfer failures can arise from any of these differences.
  • Exact type mapping is conservative but not proof of biological homology. Potential sexual dimorphism remains relevant even for mapped types.
  • Connectivity-only LIF omits many biophysical, neuromodulatory, sensory, and state-dependent mechanisms. A nonzero MN9 response is a model prediction, not observed movement.
  • The positive rule is any MN9 rate above zero. This protocol does not estimate a calibrated behavioral probability or effect size.
  • Ten repetitions characterize simulation-seed variability, not animal-to-animal biological variance.
  • No claim of consciousness, sentience, understanding, whole-animal fidelity, or clinical relevance is made.

Sources and provenance

Supplementary Table 3 workbookSpringer Nature supplementary workbook · SHA-256 6922e16825aa0c92a28e2a634b073dcad7641e9c9283c4b9d8ee1e34e6d9b8d9
Shiu identity resourcePublished source · converted JSON SHA-256 7f35a045076c2360b498636a73408618e8b5492e148ea777e508fe6cc24e8668
FlyWire annotations v3.1.0Versioned annotation table · SHA-256 9a4f8b2f843196074431ebd7cd883536afa1be86c8a4ce90970441e8be81d1be
MaleCNS v1.0 body annotationsSHA-256 2177e246113e4cfbf1e7772ec37c6da1955ff22e8063d0b1f833101f99a9a3b2