Work

Usnica 4.1 GPU: A Multiscale Bayesian-Calibrated Framework for Whole-Brain Simulation at 86 Billion Neurons on a Single Consumer GPU

Year: 2026
Type: preprint
Abstract: Alzheimer's disease affects 55 million people worldwide. We present Usnica 4.1 GPU, a multiscale Bayesian-calibrated framework integrating Knowles-Cohen amyloid kinetics, Hodgkin-Huxley neuronal dynamics, and Tononi spectral complexity for whole-brain simulation. The complete system couples equations across molecular (amyloid-beta 42), cellular (membrane, calcium, ATP), and network scales. Direct integration at short time steps with 86 billion neurons requires 3.44 TB of VRAM and 19.1 hours per second of biological time on an RTX 4060, rendering 50-year trajectories infeasible. We introduce three equivalent approximations: an Izhikevich surrogate with less than 0.12% peak error, a voxel mean-field at an 860,000-to-1 ratio, and 10,000-probe spectral complexity. The fast model reproduces lag time (18.32 vs 18.45 hours) and peak potential (115.10 vs 115.24 mV) while reducing memory to 0.016 GB and execution time to 2.7 seconds for 86 billion neurons across 0.5 biological seconds. In 10 minutes of GPU execution, we validate 10 compounds against trial plaque reduction with a mean absolute percentage error of 8.4% and propose the Usnica-Safe GRAS triple combination (inhibiting nucleation, reducing calcium, and boosting ATP) with a 40.9% reduction in toxic oligomers, safe for long-term use over 40 years. Neonatal calcium and ATP variations predict a 20-year onset shift over an 80-year lifespan.
Language: English
Open Access status: green
OpenAlex ID: W7204201427
Cites: 35
Cited by:
Related to: 10