The depth of the shared Kenyon-cell mode determines olfactory learnability: diagnosis and cure of connectome-specific learning in a whole-brain spiking model

Abstract

Connectome-constrained whole-brain models of Drosophila reproduce sensorimotor responses without per-cell tuning, yet associative learning inside a full spiking recurrent network remains an unsolved benchmark. We show that olfactory learnability is a property not of the rule and not of the neuron, but of the connectome, set by how far real odour codes escape the shared, attractor-dominated Kenyon-cell (KC) mode. On MaleCNS v1.0 (166,700 neurons; 25,582,938 edges), KC codes for DM1/VA1v lie at Jaccard J = 0.25 versus an excitability null of J = 0.53 +/- 0.06, escaping the mode, and the pair learns after a single presentation (one-shot). On FlyWire v783 the same pair lies at J = 0.76 versus a null of 0.896 +/- 0.02, inside the mode, and learning loses specificity. The root cause is dynamic: on v783 the olfactory drive is weaker (ORN-to-PN at x0.46/x0.15 of MaleCNS), the KC module is weaker (x0.39), and APL is weak relative to drive (0.27 vs 0.45). A pre-registered cure (F2-heal; sha256 e80f28bf) strengthens the ORN-to-PN input (x4) at APL 5.8, drives J down to 0.501, and restores one-shot odour-specific learning (SPEC_exc 2.16, 7/8 seeds >= 1.25, no-reward zero, directed swap). On Usnea we show that a neuron's sign is a wiring property: the official v783 sign is excitatory (84 edges) while the annotation says GABAergic, yet in our network the sign is inert because Usnea lies on no water-to-MN9 path. Learnability is a property of the data, not of the substrate; the benchmark measures the wiring, not the model. Project channel: "Правила игры" (Law of the Game) - https://t.me/law_of_the_game - project notes and popular-science writeups of this series.