Fly Rig Anatomy

With the connectome's own synapse strengths and every cell of undecided transmitter signed excitatory, the resting firing of the olfactory receptor neurons alone drives the projection neurons to more than ten times their resting rate and the Kenyon cells to their ceiling. The afferent synapse is the fly's; the excess is recurrent excitation inside the lobe, and the 67 local neurons with no transmitter prediction carry it: signed by their class's majority, inhibitory, the lobe rests at the fly's rates.

Measured The antennal-lobe gain experiment: the 2,639 receptor neurons receive random synaptic events calibrated so that an isolated cell fires 8 Hz⁴⁵, nothing else is driven, 10 s. The antennal lobe reaches its plateau inside 40 to 60 ms and the mushroom body follows within 100 ms; in the last two seconds the receptor neurons fire 21 Hz (the lobe feeds back on them), the projection neurons 78, the local neurons 140, the Kenyon cells 92, APL 399, the whole network 6.1 Hz with about 4,000 cells at or above 100 Hz. The gain from receptor to projection neuron is 9.7 against 0.6 in the fly, from projection neuron to Kenyon cell 1.2 against 0.02²⁶. The graph's receptor-to-projection connection is the fly's rested unitary synapse (14.9 contacts, 4.6 mV, 49 partners per projection neuron against a measured 6.2 mV and about 40 partners³⁵), so 49 afferents at 8 Hz put a projection neuron above threshold by themselves, 9 mV of mean drive against a 7 mV gap; the projection neurons then give each other 21 mV, the local-neuron mass is net inhibitory onto them by only 10,000 mV of 100,000, with 22,700 mV of excitation resting on cells with no predicted transmitter, and each Kenyon cell receives 40 mV from other Kenyon cells against 30 taken by APL. Giving the afferent synapses the short-term depression Kazama and Wilson measured (release probability 0.79, recovery 120 ms derived from their 40 % depression at 7 Hz) halves the excess (projection neurons 59 Hz, Kenyon cells 47, network 4.3, the afferents back at 9.7 Hz), and at Shiu's 0.275 mV keeps the mushroom body dark for 10 s while the lobe holds 54 Hz; no depression setting gives a state between silence and the attractor. The antennal-lobe sign-policy experiment then ran the same input under four policies for the 3,177 cells with no transmitter prediction. Signed by the predicted majority of their annotation class, which negates exactly the 67 antennal-lobe local neurons (their class is predicted 214 inhibitory to 137 excitatory), the lobe rests: projection neurons 4.3 Hz, within the fly's 2 to 5, a receptor-to-projection gain of 0.5 against the fly's 0.6, Kenyon cells 0.4 Hz, the network at 1.6 Hz with only the ring loop of Figure 4 at the ceiling. Silencing the 67 alone gives the same (projection neurons 5.1 Hz); the curated transmitters Virtual Fly Brain holds for 15 of them change little (projection neurons 71 Hz against 78); signing every unpredicted cell zero breaks the feeding pathway (MN9 at 12 % of its plateau, because the unpredicted cells include 268 gustatory afferents). With the Kenyon cells at their measured threshold gap as well, the live model, the Kenyon cells fall to 0.00 Hz.

Bounded by A class statistic applied to types whose own predictions are cholinergic (three of the 19 types, about 4,000 of the 71,600 mV the 67 cells put out); the receptor neurons fall to 1.7 Hz under the class-majority sign, below their 8 Hz, because the local neurons' inhibition of the afferent terminals, presynaptic and divisive in the fly⁴⁶, is whole-cell inhibition in a point neuron; the lLN1_bc sign (finding 5); one seed of the background events.

Nearest in the literature The synapse and its depression³⁵, the presynaptic lateral inhibition whose strength at rest no source gives⁴⁶, the all-to-all APL feedback⁴⁷, and the argument that a fifth of the local neurons may be cholinergic¹⁸ are all measured on the fly, not on a model. Shiu et al. start every cell at 0 Hz and report no antennal-lobe rate²; a hemibrain olfactory spiking model keeps its Kenyon cells sparse only with a hand-tuned projection-to-Kenyon scale and homeostatic scaling in the projection neurons⁴⁸; a MaleCNS GPU simulator records that its glomerular selectivity rested on local interneurons spiking⁷⁴. The resting gain of a whole-CNS model at the connectome's own weights was not found.

To reproduce Random excitatory events of 4.85 contacts on the olfactory receptor neurons only, 840 per second per cell, all else silent, 10 s at 0.31 mV with lLN1_bc GABAergic and adaptation off; compare the last-two-second rates of the receptor, projection, local, Kenyon and APL populations; repeat with the afferent depression on, at 0.275 mV, and with the 67 unpredicted antennal-lobe local neurons signed inhibitory or silenced.

Spontaneous synaptic input per cell class, at the rates the literature reports, does not give the model a resting state: at the calibrated level it ignites an attractor within a tenth of a second, the olfactory one when undecided transmitters are signed excitatory and the ring loop under the live model, at a little over half that level it is silent, and there is no level in between.

Measured The background-activity experiment: 27 cell classes, each cell its own random train of 1.5 mV events (4.85 contacts, the graph's mean connection) at a rate found on an isolated cell (olfactory receptors 8 Hz, compass neurons 4⁴⁰, other central cells 1 to 5; projection neurons, Kenyon cells, descending and motor neurons nothing, since their rest is inherited^26,35), 60 s. At the calibrated table the antennal-lobe / mushroom-body attractor ignites in 60 to 80 ms and holds for 60 s (network 7.1 to 7.6 Hz, 4,149 cells at or above 100 Hz, Kenyon cells 92, projection neurons 77; one of 14 judged classes in range; Fano factor of the population count 11.7; a third of the active cells firing at their refractory interval). At 0.7 of the table the same state after 180 to 220 ms; at 0.5 and below, 23 spikes in 20 s from all 166,700 cells. The receptor events alone ignite the olfactory loop; central events alone ignite the compass ring. On the background the escape response stands out (Giant Fiber 293 Hz), the steering response is marginal (DNa02 44.5 Hz against a 5 Hz background with a standard deviation of 18), and the feeding and small-object responses are lost (MN9 2.5 Hz against 86.5 without the background; no DNp09 spike). With the afferent depression of finding 1 on, two classes are in range instead of one and the same responses are lost. Those runs signed every cell of undecided transmitter excitatory. Under the live model the same background ignites the ring loop of Figure 4 instead, from the start (network 2.8 Hz, 522 cells at or above 100 Hz, the Kenyon cells silent, projection neurons 2.3 Hz, receptor neurons 3.7); the lower levels, run with GLNO signed excitatory, give the loop at 0.7 of the table and silence at 0.5, so there is still no level in between; on the background the escape response still stands out (Giant Fiber 297 Hz), the steering response is marginal (DNa02 31 Hz against a background of 6 with a standard deviation of 20), and the feeding and small-object responses are still lost (MN9 1.5 Hz, no DNp09 spike).

Bounded by Excitatory events only; the lLN1_bc sign (the predicted sign gives the full attractor at 10.3 Hz); per-class adaptation on top turns the plateau into bursts (Fano factor 482); the two kernels agree in distribution, not spike for spike, because they draw different random streams.

Nearest in the literature Shiu et al. assume 0 Hz everywhere²; a resting-state fly model is a named open gap⁴³; the asynchronous irregular state is a balanced-input result from cortex^33,49. Fly64 adds background drive and noise to every neuron and reports no resting-state measurement⁶⁰. A MaleCNS port of the Shiu model reports that a uniform tonic baseline destroys stimulus specificity and abolishes the escape response and names cell-specific spontaneous activity as the missing ingredient⁷³. Li et al. fit a FlyWire rate model to whole-brain calcium recordings and find a sparse set of inhibitory hub neurons sustaining the resting state⁵⁰, one of them an antennal-lobe local neuron type, lLN2F_b, that sits at the ceiling of this model's olfactory attractor. A spiking whole-CNS model given class-specific resting input, and what it does with it, was not found.

To reproduce The class table at 1.0, 0.7 and 0.5 of its rates, 60 s from silence; compare the late-window rate per class, the count of cells at or above 100 Hz, the population Fano factor, and the four reference pulses (loom, one-eyed motion, sugar, small object) against the same pulses without the background.

Under the live model every effective input leaves the network in the same self-sustaining state, the ring loop of Figure 4. The antennal-lobe / mushroom-body loop that taste and most lobula pulses ignited while every cell of undecided transmitter was signed excitatory is removed by the class-majority sign of finding 1; the Kenyon cells' measured threshold gap alone keeps the mushroom body out of it but leaves an antennal-lobe loop standing; and that olfactory loop, not the stimulus, carried most of the forward drive the robot's throttle read after a pulse.

Measured The latch, sign, taste, small-object and loom experiments re-run under the live model in runs of 4 to 8 s. Every pulse ends in the ring loop with the Kenyon cells at 0.0 Hz in every 20 ms tick, lLN1_bc at 0 to 2 Hz and APL at 28 to 37 Hz; none releases without adaptation. Each setting alone, on the same pulses: the class-majority sign alone keeps the lobe dark under a sugar pulse (projection neurons 0.25 Hz, Kenyon cells 0.2) and puts the LC9 and LC10c pulses in the ring loop with an excursion of the mushroom body's giant cells (Kenyon cells 3 Hz, APL 258 to 262, DPM 142); the Kenyon-cell gap alone keeps the Kenyon cells at 0.2 to 0.5 Hz but leaves the lobe latched by itself (local neurons of the lLN2 types at 276 to 312 Hz, projection neurons 58, APL 156 to 162, about 1,000 cells at or above 100 Hz, the network at 2.7 Hz). That lobe-only loop holds the leg motor pools at 5.4 Hz and the throttle at 0.20 after an LC9 pulse (93 ticks of forward command), as the joint loop did (5.3 to 5.5 Hz, both wheels on for runs of up to 5.5 s); with the lobe at rest the pools fall to 2.5 to 3.2 Hz, the throttle to 0.09 to 0.14 and the runs to under 0.4 s. Under the build's excitatory sign the joint loop stood at 4.2 Hz with the Kenyon cells at 47 and 2,800 cells above 100 Hz; which pulse landed in which state was not its strength (right-eye LC9, LC10c with LC10d and the calibrated zoom in the ring, bilateral LC9, LC11, LC4 alone and LPLC2 alone in the olfactory loop), and of 53 lobula types driven alone every effective one ignited one or the other within 20 to 200 ms.

Bounded by The common conditions; runs of 4 to 8 s, so a later release is not excluded; the throttle is the decoder's baseline-subtracted leg-pool reading, so the comparison is of the ignition hum, not of walking.

Nearest in the literature DOOMFLY's notes that its network stays busy⁵⁹, without a characterisation; Shiu et al. and Sapkal et al. run one-second trials from 0 Hz and report no persistent state^2,39; the Loihi 2 port of the FlyWire model scaled its synapses down to negligible before adding a background, to prevent runaway excitation⁴⁴; a MaleCNS port of the Shiu model divides the weight by a synapse-density factor and requires activity to die out after each stimulus, which it does⁷⁵, the regime this graph enters below about 0.25 mV (finding 4). Which state a pulse ends in, what removes the olfactory one and its share of the motor output were not found.

To reproduce A 0.5 s pulse of 15 mV on T4/T5 subtype a in both eyes, 1 s of 30 mV on the 34 sugar neurons, and 0.5 s of 15 mV on both LC9 populations, in runs of 4 to 8 s from silence, under the four combinations of the unpredicted-transmitter sign (excitatory or class majority) and the Kenyon-cell threshold (7 or 21.5 mV); rank cell types by rate in the second after the pulse and compare the leg pools and the throttle.

The one free parameter of the model, fitted at 0.275 mV on the FlyWire female brain, gives 68 % of maximal MN9 firing on MaleCNS where the criterion asks for about 80 %; the criterion is met between 0.30 and 0.32 mV.

Measured The weight calibration: 0.15 to 0.45 mV per contact, the 34 sugar neurons held at 100, 150, 200 and 300 Hz by a tonic drive re-found at each weight, 2 s, MN9 at 100 Hz against its plateau over the three higher rates. At 0.275 mV MN9 gives 60.5 Hz against a plateau of 89.2 (68 %, or 71 % against the plateau mean); 0.30 gives 77 %, 0.32 94 %, 0.33 101 %, 0.35 85 %; the scatter in the flat top moves the ratio by 10 to 15 points between neighbouring weights, so the answer is a band. Within 30 % of 0.31 the pathway keeps working in Shiu's sense (MN9 rising monotonically from 44 to 118 Hz, the proboscis pool from 26 to 51); below about 0.25 mV the model changes regime (no ignition, MN9 at a tenth of its own plateau). The sweep ran with the predicted signs; repeated at 0.31 with the drives re-found under the live model's signs and threshold, MN9 at 100 Hz is 96 % of its plateau maximum and 97 % of its mean, with the proboscis pool at 41.5 Hz.

Bounded by A regular tonic drive where Shiu used irregular Poisson trains; both hemispheres driven and both MN9 cells read where they drove one side and read the other; a two-cell readout that cannot be averaged over 30 trials; a male whole CNS against a female brain; and the olfactory loop igniting under sugar from 0.25 mV upward, the likeliest reason this graph wants more weight than FlyWire did.

Nearest in the literature Shiu et al. state the criterion and report the model robust to a 30 % change². An independent MaleCNS port of the same model chose its per-contact value by a sweep against escape and auditory criteria and reports 0.30 and 0.35 passing with 0.25 and 0.40 failing, without a feeding readout, because its build does not type the taste neurons⁷³: consistent with this band, by a different criterion. Another MaleCNS port moves the weight the other way: it divides it by a synapse-density factor, since a male neuron carries a median 1.81 times the synapses of its female counterpart, and passes its own 29 validation cases, sugar to MN9 among them, with activity dying out after each stimulus⁷⁵. The MN9 criterion itself repeated on MaleCNS was not found.

To reproduce For each candidate weight, find the tonic depolarisation that holds the 34 sugar neurons at 100, 150, 200 and 300 Hz over 2 s; compare MN9's rate at 100 Hz with the largest of the other three.

The self-sustaining olfactory state rests on the sign of thirty antennal-lobe local neurons, lLN1_bc, which the connectome predicts cholinergic and the curated literature calls GABAergic: with the predicted sign their self-excitation lights the lobe whatever sign the undecided cells of finding 1 carry, and with the curated sign the lobe rests under the live model; the sign of the DPM cell, the one the classifier's own paper flags, changes nothing that matters.

Measured The sign scan. In the graph lLN1_bc's 30 cells put +62,985 mV out, +23,888 of it onto themselves and none onto the Kenyon cells; DPM puts +9,019 mV onto 3,989 Kenyon cells; APL puts -53,955 onto 4,063. Under a 1 s sugar pulse at 0.275 mV with the predicted signs the loop stands at 7.27 Hz with the Kenyon cells at 122.6; with DPM flipped it stands (6.17 Hz, Kenyon cells 84); with lLN1_bc flipped it is gone (1.26 Hz, Kenyon cells 0.08, APL 17.5), and with both the same. Under a visual pulse lLN1_bc fires no spike, so its sign cannot matter there, and flipping DPM alone ignites the loop the network otherwise avoids (APL falls to 0, then the Kenyon cells go to 89 Hz). At 0.31 mV the curated sign leaves a sugar loop at half strength (4.28 Hz, Kenyon cells 47, lLN1_bc itself at 158 Hz with every output inhibitory, driven by the loop rather than driving it), and it is what keeps a visual pulse out of the mushroom body, which with the predicted sign at that weight ignites the full attractor 140 ms into the pulse. With the undecided local neurons signed by their class's majority, the predicted lLN1_bc sign still lights the lobe on the receptor neurons' resting input alone (projection neurons 57.5 Hz, Kenyon cells 92 without their threshold gap, the network at 5.4 Hz), where the curated sign leaves it at rest (4.3 Hz); under the live model lLN1_bc fires 0 to 2 Hz after every pulse.

Bounded by The call itself: acetylcholine at a confidence of 0.75 for all 30 cells¹, GABA as curated⁶³, with the classifier's paper arguing that lateral excitation is more common than thought and that a subset of lLN1 releases both GABA and glutamate¹⁸. The claim is that the model's state rests on the call, not that the call is wrong.

Nearest in the literature Eckstein et al. name DPM as a known misprediction and the antennal-lobe local neurons as a split hemilineage¹⁸; the MaleCNS paper gives the classifier's confusion matrix¹. The sensitivity of a whole-brain model's resting state to one cell type's sign was not found.

To reproduce The sugar pulse (30 mV, 0.2 to 1.2 s of a 4 s run) at 0.275 and at 0.31 mV, each with lLN1_bc as predicted and as curated; compare the network rate, the Kenyon-cell rate and the count of cells above 100 Hz in the second after the pulse; then the receptor neurons' 8 Hz background for 10 s with the undecided local neurons signed by their class's majority, lLN1_bc each way.

Driven at the same level, the 26 high-salt labellar neurons, aversive in the fly, drive MN9 harder than the 34 sugar neurons do, and hold it after the drive, with the olfactory circuit at rest.

Measured The taste-modality experiment: each population at 30 mV for 1 s from silence, with the class-majority sign and the Kenyon-cell gap and GLNO signed excitatory. Sugar: proboscis pool 39.8 Hz, MN9 94.5; water (17 cells): 24.9 and 20.5, a transient (31 Hz in the first half-second, none after); high salt: 41.8 and 107.5, with the left MN9 at 162 and the right at 53; bitter (38 cells): 7.4 and 4.0; the silence control exactly 0. In the 0.3 s after the drive high salt holds MN9 at 103 Hz and the proboscis pool at 43, with the Kenyon cells silent. With GLNO inhibitory, as in the live model, the sugar drive gives MN9 89.5 Hz and falls to 23 Hz within 0.3 s of its end; the high-salt run has not been repeated under that sign. Under the build's excitatory sign for undecided transmitters the same order held (sugar 86.5 Hz of MN9, high salt 107.5), and at 0.275 mV with the predicted signs as well (high salt 57.5 against sugar's 72.5).

Bounded by The common conditions and the modality typing, which is the MaleCNS taste paper's²⁰; the graph encodes connectivity, not valence, and nothing was tuned to suppress it.

Nearest in the literature Shiu et al. report that the second-order cells below the sugar neurons receive high-salt input in the fly¹⁹; Tastekin et al. type LB3d as high-salt avoidance neurons and compute effective connectivity from all four labellar subtypes to MN9 and the other feeding motor neurons within a few hops²⁰, a structural measure; the Shiu model was run with sugar, water, bitter and Ir94e neurons, not high salt²; the MaleCNS port that drove all labellar afferents together got no feeding output⁷³; another MaleCNS port reports that six of the neurons in Shiu's female sugar set are LB3d-like, the subtype matched to high-salt lines, and supply about a quarter of the MN9 response there⁷⁵, the nearest result and one from inside the sugar set. The high-salt population driven on its own was not found.

To reproduce Drive each of the four labellar populations at 30 mV for 1 s in a 1.5 s run from silence; compare the proboscis pool and MN9 during the drive and in the 0.3 s after, and the Kenyon-cell rate after.

Under its documented input, DNp09 fires once per cell at the onset of the object and never again, because the pulse ignites the network and the network's ascending neurons put CL366, DNp09's strongest inhibitor, at its ceiling for the rest of the run.

Measured The LC9 forward experiment and the DNp09 trace. LC9 at 15 mV in both eyes for 0.5 or 2 s: two DNp09 spikes, one per cell, in the first 20 ms tick and none for the rest of the pulse or the 1.5 s after; LC9 itself falls from 50 Hz in the first tick to 5 or 6 Hz; CL366 (two cells, GABAergic as curated and as predicted⁶³) sits at 97 to 118 Hz during the pulse and 102 to 114 after, from the second tick onward; LC9's own feed-forward inhibitors PVLP005 and PVLP020 stay at 2 to 13 Hz. In the graph LC9 is DNp09's largest visual input on both sides (1,130 and 847 contacts), CL366 delivers -47 to -49 mV per spike onto the left cell and -15 to -19 onto the right, and under sustained LC10c drive at 15, 30 or 45 mV DNp09 receives two to three times more inhibition than excitation. DNp09's own output onto the leg pools is 0.8 and 5.8 mV per spike against 166 and 164 for DNa02, so even a sustained DNp09 would need 240 to 510 Hz to move a leg motor neuron directly; the forward command the decoder emits 80 ms after the onset is the ignition of finding 3. Under the live model the same two spikes fall in the same tick. With per-class adaptation on, DNp09 fires through the pulse instead, 12 spikes over its half-second, with CL366 at 41 Hz, off its ceiling, and LC9 holding 14 to 15 Hz; the benchmark, under the live signs, gives 15 (finding 14).

Bounded by A harness pulse, not a real object; the network state that silences DNp09 is the attractor, so a resting state (question 1 of section 10) could change the answer, and adaptation, which weakens that state, already does; one graph.

Nearest in the literature Bidaye et al. find P9 silent during spontaneous walking, an initiator of pursuit, with LC9 as its direct input¹⁴; Sapkal et al. show the halting neurons FG and BB inhibiting the descending neurons P9 recruits, in flies and in the Shiu model with P9 stimulated directly³⁹. Neither reports P9 as an onset detector under sensory drive, nor CL366 as its inhibitor; Virtual Fly Brain has a transmitter for CL366 and no function⁶³. Not found.

To reproduce 15 mV on both LC9 populations from 2 to 2.5 s of an 8 s run from silence; compare DNp09's spike times, LC9's rate per 20 ms, CL366's rate during and after, and the leg pools.

The cells the robot's gyro used to drive reach the wing motor pool seventeen times more strongly than the leg pool, have no direct connection to DNa02, and reach the DNa02 on their own side with a net negative sign two synapses out; the engineered input had no wiring behind it.

Measured The proprioception review, on the built graph: per source cell at one hop, a left haltere afferent puts +5.2 mV onto the left wing motor pool and +0.3 onto the left leg pool (unsigned three-hop influence 0.15 against 0.014); no monosynaptic edge onto either DNa02; at two hops -30.2 mV onto the left DNa02 from the left population and -53.1 onto the right from the right, against +2.4 mV per leg chordotonal cell onto its own leg pool at one hop. In the closed-loop experiment the gyro-to-haltere path raised the contralateral DNa02 from 6 to 13 Hz as intended, and with it depressed the leg motor neurons to 1.7 and 1.5 Hz for the second after the stimulus (6.2 and 5.9 without it), never settled (chordotonal cells at 78 to 89 Hz and the command still on at 6 s) and locked the two leg pools together (correlation 0.80 against 0.44). On the floor the assignment that had worked better (83 % direction-following against 67) had been found by trial.

Bounded by A structural measurement under the sign policy, at one and two hops; the influence numbers compare paths from one source, not across sources; the new default has not been re-measured on the floor.

Nearest in the literature The halteres' reflexes act on wing kinematics⁵, their afferents reach a wing steering motor neuron in the blowfly, and a walking fruit fly holds them still⁶: the wing route is the physiology. The signed route from the haltere afferents to the walking steering neuron in a connectome was not found.

To reproduce From the built graph, sum the signed weight per source cell from each haltere population onto the wing pools, the leg pools and DNa02 at one and two hops; run the closed-loop experiment with the gyro into the halteres and without.

In this connectome the courtship-tracking LC10 family reaches the right pursuit neuron and not the left, while the documented input LC9 reaches both.

Measured The DNp09 trace: onto the right DNp09, LC10c 29 cells and 194 contacts, LC10a 16 cells and 65, LC10d 6 cells and 44; onto the left DNp09, none from any LC10 type, and no relay through the anterior optic tubercle carries it; LC9 1,130 contacts onto the left and 847 onto the right. Virtual Fly Brain's male connectome gives the same picture: 18 LC10c cells onto the right DNp09 at 5 to 13 synapses each and none onto the left, LC9 onto both⁶³. In the harness a right-eye LC10c pulse gives DNp09 as many onset spikes (two) as a bilateral one.

Bounded by One male, its proofreading, and connections of 5 to 13 synapses per cell, where a single missed or false edge counts; a build at the connectome's 0.5 confidence threshold.

Nearest in the literature Bidaye et al. found that LC10 activation did not drive P9 and that LC9 does¹⁴; the LC10 family tracks the courtship target through LC10a and the anterior optic tubercle¹²; the MaleCNS paper reports LC10a isomorphic between the sexes, its targets dimorphic, and the optic-lobe connections of the male strongly bilateral¹. A one-sided LC10 input to DNp09 was not found.

To reproduce From the public MaleCNS synapse table, list DNp09's presynaptic cells per side and group them by type; drive LC10c on one side, then the other, at 15 mV.

With the sugar neurons of both sides driven alike, the left MN9 fires three to seven times the right, at every weight and sign tried; driven from one side only, the left cell still dominates.

Measured The taste-modality experiment and the sign scan, from per-cell counts: during 1 s of bilateral sugar drive the left and right MN9 fire 142 and 37 Hz under the live model, 150 and 39 with GLNO excitatory, 135 and 38 under the build's excitatory sign for undecided transmitters, 128 and 17 at 0.275 mV with the predicted signs, 139 and 31 with the curated sign at 0.275, 149 and 35 with the predicted sign at 0.31; under high salt 162 and 53, water 113 and 29, bitter 12 and 0; in the calibration protocol at 100 Hz, 109 and 12. The ratio narrows from 7.5 to 3.6 as the weight rises because the right cell gains more. The feeding experiment of Shiu's validation (finding 12) then drove one side at a time at 100 Hz: the right sugar neurons give the left and right MN9 148 and 38 Hz, the contralateral cell ahead as Shiu found, and the left sugar neurons 143 and 44, the ipsilateral cell ahead; bitter or Ir94e neurons driven on one side drive the left MN9 to 138 Hz where driving both sides gives 1.5 to 4.5 Hz.

Bounded by Two cells in a deterministic model; a tonic drive where Shiu used Poisson trains; one graph.

Nearest in the literature Shiu et al. report unilateral sugar input driving the contralateral MN9 harder than the ipsilateral one²; Tastekin et al. describe the feeding motor neurons as one or two left-right pairs²⁰. An MN9 pair whose left cell dominates whichever side is driven was not found; whether it is the male's wiring, its proofreading or a sign is open.

To reproduce 30 mV on the 34 sugar neurons for 1 s from silence; compare the two MN9 cells' rates during and after; then hold one side's sugar neurons at 100 Hz, then the other's, and read both cells.

Shiu's policy signs the dopaminergic cells excitatory at the model's 5 ms synapse. With the antennal lobe lit, their output is what brings the Kenyon cells into the olfactory loop; with the lobe at rest and the Kenyon cells at their measured gap, it is what holds APL, DPM and the output neurons above silence; and removing it makes the Kenyon cells busier rather than quieter.

Measured The mushroom-body rest experiment: the receptor neurons on their 8 Hz background, nothing else driven, 10 s, with the output of the 392 dopamine cells set to zero or left as signed, rates over the last 2 s. In the graph the dopaminergic cells put +17 mV onto each Kenyon cell per volley (32 partners at 0.54 mV), +1,255 mV onto APL and +127 onto each output neuron. With the undecided transmitters signed excitatory and the lobe at its plateau, removing them takes the Kenyon cells from 92 to 8.2 Hz and APL from 399 to 310. With the class-majority sign and the Kenyon cells at the model's 7 mV gap, it takes the Kenyon cells from 0.4 to 1.1 Hz (184 to 348 of 4,064 active), with APL at 67 to 71 Hz, DPM from 46.5 to 33 and the output neurons from 24 to 13; silencing all 541 monoaminergic cells gives the same picture. With the Kenyon cells at Turner's gap as well, as in the live model, and silent, removing them takes APL from 36 Hz to 0, DPM from 42.5 to 0.2 and the output neurons from 21 to 9.

Bounded by A fast excitatory synapse is the policy's form, not the fly's; the fly's resting rates of APL, DPM and the output neurons are not in the sources consulted, so the numbers above set no target; why the Kenyon cells rise is not isolated (APL's rate holds, while the lobe shifts at the same time, projection neurons 4.3 to 2.3 Hz and receptor neurons 1.7 to 3.6); one seed.

Nearest in the literature In the fly the dopaminergic cells act on the Kenyon-cell to output-neuron synapses of their own compartment through D1-like receptors⁵¹, depress them when paired with odour⁵², act through distinct receptor pathways with distinct timing⁵³, and evoke in an output neuron a slow depolarisation that peaks more than 2 s after a brief stimulation, largely by volume release⁵⁴. Shiu et al. assign the monoamines to the excitatory category and note that they are modelled less well²; a MaleCNS port gives them no fast effect at all⁷⁵. What the fast-excitatory policy does to the mushroom body of a whole-CNS model at rest was not found.

To reproduce The receptor neurons' 8 Hz background for 10 s under the four combinations of the unpredicted-transmitter sign (excitatory or class majority) and the Kenyon-cell threshold (7 or 21.5 mV), each with the dopamine cells' output at zero and as signed; compare the Kenyon cells, APL, DPM and the output neurons over the last 2 s.

Carried to the male connectome by cell type and run as Shiu ran it, 120 of the 164 experimental predictions his FlyWire model was scored on hold here, against his 150, and the benchmark, scoring the same runs under a protocol adopted for adapting cells, counts 130: the whole-circuit predictions hold, and the shortfall is in which single cells the feeding response needs and which cells extend the proboscis on their own.

Measured The feeding, activation, silencing and grooming experiments and their scorecard, under the live model, each driven cell held at its target rate by a tonic drive, one deterministic run per row, each run recording the full model it was built with. Cell types were carried over through the FlyWire type annotations: 99 of Shiu's 106 feeding types map (33 by Shiu's own name in the male connectome's synonyms, 66 through a FlyWire type), 7 have no cells here, and 120 of 164 predictions hold (77 % of the 156 the graph can be asked). The whole-circuit rows hold: 12 of 14 sugar-responsive cells, 8 of 10 water-responsive cells, all four bitter and Ir94e suppression predictions (on sugar at 100 Hz, bitter or Ir94e at 100 Hz take MN9 from 93.5 Hz to 4.5 and 1.5; at 200 Hz bitter eliminates it and Ir94e leaves 15 %), the three antennal mechanosensor responses, and silencing the grooming interneuron aBN1 takes both grooming descending neurons to 0. The shortfall: 3 of 10 sugar-required and 6 of 11 water-required cells (Shiu 6 and 10), because silencing any one second-order sugar cell moves MN9 by 0 to 10 % here and the premotor Roundup by 52 %, where his model had Clavicle and Zorro carrying 33 and 23 %; and 79 of 106 sufficiency rows (Shiu 101), five of the 20 misses being one to three spikes. The benchmark now scores Shiu's rows under a protocol adopted so that a model with adaptation can be scored on the same terms: the receptor-neuron rows hold the driven cells at their target rate, the sufficiency rows keep the drive found on the model without adaptation, and a sufficiency row counts MN9 as driven at 5 % of the model's own strong-sugar response instead of above 0 Hz. Under it the live model scores 130 of 164, with 89 of the 106 sufficiency rows; this is a change of scoring, not of the model, and the count under Shiu's own criterion, 120, stands beside it. Under a water drive the proboscis motor program is all or none within the second: MN9 and about 30 gnathal cell types either fire to the end of the drive or stop together within 100 ms, so a water-silencing row records whether the silencing changed that switch, not the drive the cell carries; read against the program while it runs, only Roundup's silencing lowers MN9. Signing GLNO inhibitory moved ten rows against the run before it (124 of 164): G2N-1's requirement, which sat on the 20 % line; four sufficiency rows that now draw one to eight spikes of MN9, with the GLNO module of the ring loop no longer igniting during the drive; four water rows through the switch; and the second grooming interneuron's requirement, which rests on a drop-out of the descending neuron while the first interneuron still fires, not on its own input. Under per-class adaptation, with every row run under that protocol, 125 hold (123 under Shiu's own criterion), the bitter and Ir94e suppression at 100 Hz lost (finding 14).

Bounded by Typing across connectomes (66 types matched only through a FlyWire type; the aBN2 grooming interneuron a structural candidate); a male whole CNS against a female brain; both sides driven with a tonic drive and one run per row where Shiu drove one side with Poisson trains over 30 trials; the weight at 0.31, which puts MN9 on a plateau from 50 Hz of sugar input and leaves little room for a single silenced cell to show; the water rows read a switch whose trigger is not identified.

Nearest in the literature Shiu et al. score their own model at 150 of 164 and report 90 to 95 % of its predictions unchanged at weights 30 % lower or higher². Two MaleCNS ports hold the Shiu model to criteria of their own: 29 cases including sugar to MN9 and bitter leaving it silent, passed at a weight divided by a synapse-density factor⁷⁵, and escape and auditory criteria for the weight sweep⁷³. The 164-prediction table run on a second connectome was not found.

To reproduce Map Shiu's FlyWire cells to male types through the FlyWire type annotations; for each row of his prediction table drive the named population at his rate, or silence the named cell by zeroing its outputs, and read MN9, the named feeding cells or the grooming descending neurons against his criteria.

The ring loop of Figure 4 is five independent modules, each held by excitation between a few cells strong enough that one or two spikes of a partner carry a cell back over threshold. Nothing in a current-based unit without adaptation stops such a pair below its refractory ceiling; the octopaminergic cells that fire first hold nothing, the inhibitory members are a third of the excitation they face, and two of the five modules exist only because a policy signs an undecided cell excitatory.

Measured The ring-loop experiment (115 conditions, 4 s, visual and sugar pulses) and the graph. One spike through c contacts is a jump of 0.31 c mV that peaks at 0.157 of its size, so a connection of 144 contacts carries its target over the 7 mV gap alone; 69 excitatory reciprocal pairs in the graph clear that bar in both directions and 493 clear half of it. The modules: the two DNg33 cells, 773 and 723 contacts each way, one spike a 35 mV peak, still firing 85 to 94 Hz from their other inputs with their own outputs silenced; FR1, 18 fan-shaped-body cells all-to-all within each side, which give each other 125 of their 131 mV of drive; the compass's GLNO and PEN_a cells, +152 and +186 mV per volley onto each other against 79 and 32 from their glutamatergic partners ExR6 and ExR4; GNG016 with the taste afferents it contacts; and the abdominal pair INXXX137 and INXXX181. Silencing the octopaminergic cells that ignite first, or setting all 101 octopamine cells' output to zero, leaves every module after a visual pulse; silencing the inhibitory members raises the loop (ExR6 and ExR4 together, 440 to 580 cells at or above 100 Hz). GLNO at its measured inhibitory sign removes the compass module, GNG016 at the GABA sign Virtual Fly Brain's class carries removes the taste module after a visual pulse; no sourced sign releases the rest; per-class adaptation releases every module after both pulses in the benchmark, the DNg33 pair falling from about 120 Hz in the second after a visual pulse to 78 Hz in the last second of the run (finding 14).

Bounded by The one-spike criterion ignores summation; 4 s runs; no published threshold, time constant or rate exists for GLNO, the DNg33 pair, FR1 or the gnathal and nerve-cord types; the GLNO measurement is a population functional-connectivity result, possibly polysynaptic, by its authors' caveat²¹.

Nearest in the literature In a spiking model of the protocerebral bridge built from light-microscopy connectivity, strengthening excitation or weakening inhibition drives the compass ring into circuit-wide seizure⁵⁵: the compass module's mechanism, in one circuit with weights set by hand, where here it is one of several modules at the connectome's own weights. Li et al. find a fitted FlyWire rate model's resting state held by inhibitory hubs locked to excitatory partners, whose silencing releases the partner into runaway activity⁵⁰; here the inhibitory members exist but are a third of what they face, and silencing them raises the loop. The fly's compass holds a bump with its PEN1 cells at about 4 Hz⁴⁰, where the model with GLNO excitatory holds the whole ring at 200 Hz. A census of self-sustaining reciprocal pairs in a whole-CNS spiking model was not found.

To reproduce From the built graph, list the excitatory reciprocal pairs whose single-spike peak (0.157 times 0.31 mV times the contact count) clears the 7 mV gap each way; run the T4/T5 and sugar pulses, silence each module's members in turn, and repeat with GLNO and GNG016 at their inhibitory signs and with per-class adaptation.

With the per-class adaptation of section 5 switched on, the benchmark finds no module of the ring loop latched after either test pulse and almost no cell at the refractory ceiling on the calibrated background; the compass neurons rest inside their measured band, more cell classes sit near their target rates, and the two eyes' motion responses mirror each other. The price is the feeding motor neuron at about a third of its rate at every sugar level, a lower Giant Fiber response to looming, and a population that bursts rather than rests. It is the model's main candidate for a resting state, not the live model.

Measured The benchmark, the project's fixed battery of Shiu's validation, the four validated responses and the resting-state measures, run with adaptation off, as in the live model, and on, everything else equal. The background of the resting-state conditions specifies each class's resting rate; an adapting isolated cell fires below its target on the event rates derived for a non-adapting one, so those rates are re-derived for the adapting cell by the same procedure, never fitted to the network. The pulses run without a background. One seed, and three seeds for the background conditions. After a 0.5 s visual pulse and a 1 s sugar pulse, no module is latched in the last second, against three and four of five without adaptation, and no cell is at or above 100 Hz, against 304 and 391; the network's mean rate falls below 0.1 Hz for good 2.8 and 1.5 s after the pulse, where without adaptation it never does within the run; after the visual pulse the DNg33 pair, at about 120 Hz in the second after it, is at 78 Hz in the last. On the calibrated background 2, 2 and 4 cells sit at the ceiling at the three seeds, against 543, 450 and 476, and one, one and two modules are latched, against four, three and three: the ones left are two-cell modules just over the 100 Hz line (the GNG016 pair at 112 Hz); on the olfactory receptor neurons' background alone, none at the first seed and at most two cells and one module at the others. The P-EN1 compass neurons rest at 4.9, 5.8 and 4.9 Hz, inside the band of 1.3 to 6.5 Hz that loose-patch recordings give for standing flies⁴⁰, against 0.35, 0.25 and 0.33 Hz without adaptation, and 8 of the 14 background classes with a target rate are within a factor of two of it, against 2 or 3. Motion in one eye now gives mirrored steering: DNa02 left minus right is +18 Hz for the left eye and -23.5 for the right, against +30.5 and +2.5 without adaptation, and symmetric motion gives -9 Hz against +40.5 (question 6). DNp09 fires 15 spikes through the LC9 pulse, against 2 (finding 7). The costs: with the sugar receptor neurons held at 100, 150, 200 and 300 Hz MN9 fires 34, 25, 28 and 34 Hz, against 95 to 99, so Shiu's criterion², MN9 at 100 Hz as a fraction of its maximum at the higher rates, passes (0.99, against 0.96) only as a ratio on a flat, smaller response; at the drives that give 100 Hz without adaptation, where the adapting receptor neurons reach 60 Hz, the ratio is 0.47. The Giant Fiber fires 232 Hz to the calibrated zoom, against 302, and the other looming descending neurons a fifth to a third less. On the background the Fano factor of the population's spike count per 20 ms tick is 261, 261 and 227, against 46, 31 and 27: activity comes in network-wide bursts, where an asynchronous irregular rest would sit near 1. Shiu's validation holds at 125 of 164, against 130 (123 against 120 under his own MN9 criterion), with bitter and Ir94e at 100 Hz no longer suppressing the sugar response.

Bounded by The adaptation strengths are measured for some classes and inferred by class for the rest (section 5). Vision enters at the direction-selective and lobula cells, bypassing the photoreceptors, lamina and medulla, whose own adaptation the fly has, so the visual responses are judged without the periphery. Shiu's protocol was built for a model without adaptation even where its verdicts are fly experiments; the benchmark now holds the receptor neurons at their rate, drives the sufficiency rows at the non-adapting model's drive and counts MN9 against the model's own strong-sugar response (finding 12), and the gate's ratio does not ask for the response's size. One seed for everything but the backgrounds; runs of 4 to 20 s. These costs are why it stays a candidate: two validated responses are partly lost, the population bursts, and the visual periphery is not yet in the loop to be judged with it.

Nearest in the literature Shiu's model is memoryless and starts every cell at 0 Hz². A MaleCNS port of it carried an adaptive threshold, a standard spike-frequency adaptation not taken from fly measurements, to stop the whole CNS running on after a strong stimulus, and dropped it once a weight scaled by synapse density kept the network stable without it⁷⁵; here the weight is held by Shiu's criterion without adaptation (finding 4) and adaptation is set per class. A FlyWire rate model fitted to whole-brain recordings rests through inhibitory hubs rather than adaptation⁵⁰, and the asynchronous irregular state the Fano factor tests for is a balanced-input result⁴⁹. What per-class adaptation does to the resting state and the validated responses of a whole-CNS spiking model was not found.

To reproduce The benchmark's battery with per-class adaptation off and on, the background's event rates re-derived for each arm's isolated cell and Shiu's rows under the same drive protocol in both; compare the cells at or above 100 Hz and the modules latched after the pulses and on the background, the release times, P-EN1's resting rate, DNa02 under motion in each eye, MN9 with the sugar neurons held at 100 to 300 Hz, the Giant Fiber during the zoom and the population Fano factor, with the backgrounds at three seeds.