A three-step excitatory connectome path score found the same 10 of 14 positive cell types as the whole-brain Drosophila spiking model in its 106-type proboscis screen, with 2 more false positives

Shiu et al. (2024, Nature) reported that a whole-brain leaky integrate-and-fire (LIF) model of the FlyWire connectome predicted 150 of 164 experimental results. We asked which part of this accuracy needs the spiking model. We did not run new simulations. We used the published model rates and the same FlyWire v630 synapse counts and transmitter signs. The main test was the paper's optogenetic screen of 106 subesophageal cell types, in which 14 types caused proboscis (rostrum) extension. A rule that always predicts no extension was correct for 92 types (86.8%). The LIF model was correct for 101 types (10 hits, 1 false positive; Matthews correlation 0.78, 95% bootstrap interval 0.57-0.92). A weighted sum of excitatory paths of one to three steps onto motor neuron MN9, with one leave-one-out threshold, found the same 10 hits and 3 false positives (99 correct; Matthews correlation 0.70, 0.48-0.88). The two rules disagreed on 2 types only (exact McNemar p = 0.50). Both of these types have net inhibitory signed paths to MN9. The LIF rates ranked the types better (AUC 0.845 vs 0.769; paired difference 0.075, 0.023-0.133). This difference came mostly from 4 positive types whose neurons all have inhibitory (GABA or glutamate) labels. Every rule that uses these signs, including the LIF model, missed them. On the 58 types with at least one excitatory neuron, the AUCs were 0.992 and 0.965. A deterministic mean-field version of the same LIF model found 8 hits at 200 Hz drive. Across all 164 published tests, a rule that gives the more common result in each test category was correct 135 times (82.3%); 9 of the model's 15 extra correct results came from the screen. Thus the screen supports the claim that short excitatory connectome paths predict which cell types drive feeding. It gives little evidence that spiking dynamics add to a path count. Limitations: the screen has only 14 positive types, so the intervals are wide. We chose the path rule from 8 variants. The LIF result also depends on the 50 Hz drive and the left-MN9 read-out. We did not test FlyWire v783. The package contains code, per-type scores, metrics and figures.

connectome · FlyWire · Drosophila · leaky integrate-and-fire · whole-brain model · model validation · baseline comparison · proboscis extension · feeding circuit · MN9 · split-GAL4 screen · neurotransmitter sign · path analysis · Matthews correlation · McNemar test