Energy-efficient information processing and eligibility-trace plasticity in the Drosophila optic lobe connectome

Sensory circuits operate under finite metabolic budgets, so their information processing should be evaluated together with energetic cost. Using structural graph models, connectome-constrained dynamics, and an explicit energy proxy, we quantify energy-information trade-offs in the Drosophila optic lobe connectome. Cell-type symmetries compress connectome description length by 9–16% relative to independent-edge baselines, showing that a substantial fraction of the wiring diagram is statistically regular rather than idiosyncratic. In a canonical four-way motion-decoding regime, the real connectome supports above-chance decoding and a positive decoder-based mutual-information lower bound ($$I_{\textrm{lb}}=0.408\pm 0.050$$ bits across seeds), whereas label shuffling and connectivity shuffling collapse the bound to zero. In a separate energy-efficiency regime with topology- and weight-matched nulls, absolute decoder bounds depend on the operating point and null family, so we focus on efficiency within that regime rather than on raw information differences across regimes. Under this comparison, the real network achieves lower total energy than the connectivity null (Real/Null_Conn $$\approx 0.80$$) and higher bits-per-energy across the tested cost-weight sweep. Across retinotopic patches and compute-matched budgets, eligibility-trace plasticity (E-prop) aligns learned wiring geometry to the connectome more efficiently than REINFORCE-style policy gradients. Additional trajectory and motif analyses show that this advantage is visible over training, while motif similarity and geometry alignment capture partly distinct aspects of biological realism. Taken together, these results support a conservative interpretation: the optic lobe connectome contains compressible structure, that structure supports decodable signal, the real graph is more energy-efficient than matched nulls within the Part D regime, and eligibility-trace learning provides a plausible route toward that regime.