Connectome Wiring Shapes Motor Lesion Phenotypes in Embodied Drosophila Locomotion
Abstract
A central controversy in computational neuroscience—recently formalized as The Digital Sphinx debate (Brunton,Abe, Hu, & Tuthill, 2026)—asserts that deep reinforcement learning (DRL) can optimize arbitrary artificial neural networktopologies to generate realistic animal locomotion. Consequently, high-level behavioral replication alone is insufficient toprove that connectome-constrained models capture genuine biological representations. Here, we present a pre-registeredadversarial lesion audit testing whether biological synaptic connectivity produces functional motor specificity thatdegree-matched random networks cannot replicate under identical embodied biomechanics. Using the Drosophilamelanogaster Janelia MaleCNS v1.0 connectome, we mapped the descending motor subcircuit (4 DN → 125 VNCInterneurons → 377 Motor Neurons) controlling 42 leg joint degrees of freedom in FlyGym MuJoCo physics. We generatedan ensemble of 20 degree-preserving bipartite null models using Markov-chain double edge swaps (0.0 degree deviation;~7% synaptic overlap) and evaluated N = 10 independently-seeded biological policy instances against these 20 nullmodels under targeted in-silico unilateral ablation of descending steering neuron DNa01. Unilateral DNa01 ablation inbiological connectome models induced a stereotyped lateralized steering turn of 38.19 ± 10.83°/s (95% bootstrap CI:[32.37, 45.22]°/s), consistent with in-vivo behavioral optogenetic perturbations of DNa01 (Rayshubskiy et al., 2025). Incontrast, identical single-neuron ablations in the 20 degree-matched rewired null models yielded an average turning rate ofonly 2.92 ± 1.25°/s (95% bootstrap CI: [2.12, 3.79]°/s), with 0 of 20 null models reproducing the turning phenotype. Thesenon-overlapping distributions confirm our pre-registered falsification criterion and demonstrate that while unconstrainedoptimization can make arbitrary network graphs walk, only the authentic biological wiring diagram preserves single-celllesion vulnerability and localized motor control.Code: https://github.com/Anishp-cell/connectome-drl
