Microtubule-Augmented Perceptron: Structured Biological Topology as an Inductive Bias for Reservoir Computing

Inspired by the intracellular structure of biological neurons, we propose a computational microtubule module that provides additional internal processing within a perceptron. We use a classical automata abstraction of the microtubule lattice geometry 13 protofilaments × 20 rings with a 3-monomer helical seam as a fixed reservoir substrate, training only a linear readout. This work uses purely classical dynamics; we explicitly exclude quantum or consciousness-related claims, as decoherence timescales make such mechanisms implausible in warm biological tissue. We evaluate whether the specific structure of the microtubule lattice provides any computational advantage over topologically matched alternatives. On a static parity task, the helical seam provides no measurable advantage over a seamless cylinder (Cliff's δ ≈ 0, p ≈ 1), while the lattice broadly outperforms random and dense topologies. On sequential MNIST where 784 pixels must be integrated one at a time the microtubule reservoir achieves 37.6% accuracy with 95% CI [36.9%, 38.4%] (25 seeds), outperforming a matched random reservoir (ESN, 23.3%, Cliff's δ = +1.0, p = 2.8 × 10⁻⁹) using only 11 trained readout parameters versus the ESN's 2,610 total parameters. A trained GRU achieves a higher mean (57.6%) but with 10× wider confidence intervals and 2,500× more trained parameters, and collapses to near-chance on some seeds due to vanishing gradients over 784 steps. Our results suggest that structured biological topology is most beneficial when temporal routing over long sequences is required, and provides no advantage for static classification tasks where the seam's orbital properties are irrelevant.