Hierarchy of wiring specificity, parallel descending-neuron screens and dimorphic circuit organization in whole-CNS Drosophila connectome simulations

Complete synapse-resolution connectomes of the adult Drosophila central nervous system (CNS) link sensory inputs, central brain circuits, descending pathways and ventral-nerve-cord (VNC) motor pools within a single graph, yet it remains unclear what granularity of synaptic wiring different computations actually require, how efficiently whole-CNS models can screen causal bottlenecks, and how to compare male and female reconstructions acquired by different electron-microscopy (EM) modalities. Here we address these questions using a GPU-accelerated leaky integrate-and-fire (LIF) simulator of the male CNS (MaleCNS v1.0; 167,565 neurons, 25.6 million directed connections) and of the female brain-and-nerve-cord connectome (BANC v888; 157,789 neurons, 3.0 million connections), paired with hierarchy-structured Maslov–Sneppen null models that preserve every neuron's exact in-degree, out-degree, outgoing synapse weights and neurotransmitter sign. First, probing three canonical circuits under identical noisy LIF dynamics reveals a three-tier hierarchy of wiring specificity within the same male CNS: mushroom-body odor expansion from eight antennal-lobe glomeruli to 4,064 Kenyon cells tolerates complete randomization of projection-to-Kenyon-cell claws (67.0% versus 59.0% eight-odor decoding), descending motor steering (PFL3 to steering descending and 815 leg motor neurons) requires hemisphere-specific cell-type connectivity (surviving a 48% type-and-side shuffle at 4.11 versus 3.98 Hz in leg motor neurons, collapsing to 0.02 Hz under type-only shuffling), and central-complex compass transmission from eight EPG wedges to 44 Delta7 and PEN glomerular instances requires sub-type glomerular identity (dropping from 1.06 to 0.33 Hz and from 0.46 to 0.20 circular cosine modulation between instance and type-and-side shuffles that differ across only 0.67% of CNS connections), although fast homogeneous synapses do not sustain an autonomous memory bump after stimulus offset (3.8% persistence). Second, closed-loop 3D walking in NeuroMechFly v2 confirms across both sexes that cell-type-by-hemisphere connectivity suffices for goal-directed steering whereas midline-mixing type-only shuffles abolish steering and region or full shuffles silence brain-wide drive below the locomotor gate. Third, executing a Common-Random-Numbers (CRN) virtual Split-GAL4 silencing screen across K = 486 parallel whole-CNS slots on a single GPU tests all 485 named descending-neuron cell types in 17.2 s against strictly non-overlapping VNC motor pools, identifying DNa06 (90% loss) and DNg75 as primary leg-motor asymmetry bottlenecks and blindly recovering pIP10 first overall for courtship wing drive (47% loss), uncorrelated with VNC output synapse count (r = 0.05–0.09). Finally, an EM threshold-recall audit shows that FIB-SEM MaleCNS retains 3.57× more connections than GridTape TEM BANC even at a matched |w| ≥ 3 cutoff, precluding direct cross-dataset firing-rate comparisons, whereas targeted rewiring of genetically annotated fru/dsx subgraphs strictly within each connectome symmetrically double-dissociates male courtship song and sensory gating and female vpoDN receptivity from shared navigation circuitry. Code & Reproducibility: https://github.com/indian-aman/drosophila-whole-cns-specificity (Release v1.0.0)