Embodied whole-CNS emulation of Drosophila melanogaster: behaviours that emerge from the connectome, and behaviours that do not

Abstract

Connectome-based simulations of Drosophila are usually evaluated pathway by pathway: a sensory population is stimulated and downstream firing rates are compared with electrophysiology. We asked instead which behaviours a complete connectome produces once it is given a body, and paired every test with an explicit control. The complete male central nervous system (MaleCNS v1.0; 165,122 neurons, 25.6 million connections, brain and ventral nerve cord) was run as a leaky integrate-and-fire network coupled to a physics-based body (NeuroMechFly v2) in a world with food, obstacles, a 45 °C patch, a day–night cycle and a caged conspecific. The engine reproduces the published implementation of the same model neuron by neuron (r = 0.999). Locomotion, sugar-evoked feeding, air-puff-evoked grooming and looming-evoked escape emerge without behavioural rules. The escape is causal and specific: with condition order randomised and the intact animal measured at the start and end of every session, silencing LPLC2 and LC4 — the loom detectors of the animal — cuts the synaptic drive to the giant fibre from 33.9 to 7.2 mV (n = 12, d = −3.56, p < 0.0001), whereas silencing the same number of arbitrary optic-lobe neurons leaves it unchanged once session drift is accounted for. Every paired experiment was then repeated with the order of conditions counterbalanced, because a fixed order had already produced one false positive in this study. Three results of the original fixed-order design did not survive. The approach-versus-recede contrast in the visual system, originally the largest effect we reported (T5 d = +14.3), turned out to be dominated by the position of each stimulus in the session: the third stimulus of every session evoked the same saturated response whatever it was. Compared at matched early positions, the direction-selective channels still separate the stimuli as in the animal (T5 5.6 vs 2.3 Hz, T4 1.6 vs 5.5 Hz, p = 0.0009), but LPLC2 does not, and its selectivity is indistinguishable from that of a random group of optic-lobe neurons of the same size. The only courtship effect we had reported (pIP10, p = 0.0082) vanished (p = 0.68), and a single-animal result on sleep-promoting neurons halving a walking command neuron did not replicate. Five capacities fail with the wiring and the sensory drive intact: thermal avoidance, courtship, circadian control of locomotion, the ellipsoid-body compass and associative learning. Descending command neurons cannot be reached from internal state (141 octopaminergic neurons at 33 Hz across 87,069 targets leave DNp09 at 0.08 ± 0.10 Hz, n = 8); sub-threshold modulation works only in the presence of spontaneous release; and in the mushroom body the punishment signal arrives but the Kenyon-cell code is not sparse (67 ± 6% of cells respond to one odour, against 5–15% in the animal). A degree-preserving shuffle of the connectome abolishes the central-complex activity (EPG 5.22 vs 0.09 Hz) but not the visual contrasts, which are properties of the input rather than of the wiring. Conductance-based synapses, at a matched population rate, raise the persistence of the compass pattern (0.07 → 0.32, n = 10, p = 0.013, not significant after correction across the 32 tests of the paper) and make LPLC2 selective; the bump is a point attractor, not a ring. We conclude that a complete connectome with standard single-neuron biophysics supports sensorimotor pathways whose necessity can be shown by lesion, and does not support persistent internal states — and that in this kind of emulation the controls decide what is true: of the effects we first reported, the ones that survived were those established by lesion against a sham, not those established by a stimulus contrast in a fixed order.

Abstract of "Embodied whole-CNS emulation of Drosophila melanogaster: behaviours that emerge from the connectome, and behaviours that do not" by Francisco Javier Aldunate Rodríguez, https://doi.org/10.5281/zenodo.23004436, used under its Creative Commons Attribution license (CC-BY); markup removed, otherwise unmodified.