Benchmarking a whole-brain connectome model of Drosophila against experimental data: diagnosing knockout-prediction failures and a candidate excitatory role for the water-taste neuron Usnea

Connectome-constrained whole-brain models are increasingly used to predict which neurons control behaviour, including through in-silico knockout screens. We re-implemented the leaky integrate-and-fire (LIF) model of the adult Drosophila brain (Shiu et al., 2024; FlyWire v783) and scored it against 16 experimentally established findings with pass/fail criteria fixed before simulation. The model passed 11/16 tests; activation (stimulus-to-motor-output) tests largely passed, whereas two of three knockout tests in the feeding circuit failed at the default input strength. We identify three candidate, separable causes of knockout failure: (i) operating point: Rattle's necessity appears only near threshold (+3% at strong vs -47% at near-threshold sugar input); (ii) a structural bottleneck: Roundup carries most of the excitatory drive to the proboscis motor neuron MN9 in the model (bilateral knockout -80% to -94% in female and male connectomes), whereas acute silencing of Roundup does not reduce proboscis extension to 50 mM sucrose in food-deprived flies; and (iii) neurotransmitter (NT) uncertainty: 85 of 406 feeding-circuit neurons have low-confidence or conflicting NT predictions. For the water-taste second-order neuron Usnea, whose GABA assignment is a low-confidence prediction, removing its inhibitory output (deletion: 10 to 42 Hz) or reversing its sign (10 to 73 Hz) rescued the failing water-to-MN9 test without losing other passed tests; among 44 uncertain cell types, no other re-signing raised water-to-MN9 above 15 Hz. At an approximately matched operating point on fresh random seeds, the Usnea knockout reduced MN9 activity in the re-signed model (-26%) but not in the original (+14%), consistent with fly data; connectivity analysis shows that a GABAergic Usnea acting through disinhibition cannot be excluded. A single-neuron knockout screen of all neurons active under one simulation seed for sugar (400) and sugar + bitter (402) inputs, confirmed on independent seeds with FDR control, identified the descending neuron DNge031 as the strongest feeding brake (bilateral knockout +87% to +175% MN9). In an independent male CNS connectome, the effect was present at the calibrated synaptic scale (+55% on pre-registered fresh seeds) but weakened at lower network drive, failing our pre-registered robustness criterion. In a pre-registered expanded benchmark built from published fly experiments (104 optogenetically activated cell types; 17 silencing tests; 3 grooming tests), the re-implemented model reproduced the sufficiency screen (balanced accuracy 0.88), but the Usnea re-signing did not improve the silencing tests overall (11/17 in both models; it gained the two Usnea tests and lost two others, at unmatched operating points), and a second re-signing screen of 80 uncertain cell types in olfactory and navigation circuits found no candidate. The Usnea hypothesis therefore remains a candidate that awaits a direct test. We provide testable predictions, notably immunostaining of Usnea for cholinergic versus GABAergic identity, and release all code, pre-registration files and outputs.