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

Technical Summary: Usnica 4.1 GPU

Problem Statement
The paper addresses the computational intractability of simulating whole-brain Alzheimer's disease pathology over clinically relevant timescales (50 years) on consumer hardware. A complete multiscale model integrating amyloid-β42 (Aβ42) kinetics, Hodgkin-Huxley (HH) neuronal dynamics, and Tononi spectral complexity requires solving $3N + 5V + 4N$ equations across molecular, cellular, and network scales. For a full human brain simulation ($N = 86$ billion neurons) at a reference time step of $dt = 10$ µs, the system demands 3.44 TB of memory and approximately 19.1 hours of wall-clock time per second of biological time on an RTX 4060 GPU. This renders 50-year trajectories (requiring ~7.8 years of wall time) infeasible. Furthermore, existing frameworks lack a unified approach coupling amyloid aggregation, neuronal excitability, and consciousness metrics on consumer-grade hardware (8 GB VRAM) without ad hoc tuning.

Methodology
The author presents Usnica 4.1 GPU, a Bayesian-calibrated framework that integrates three distinct scales:

1. Molecular: Knowles-Cohen amyloid kinetics modeling monomer ($M$), oligomer ($F2$), and fibril ($F$) dynamics, with rate constants ($k_n, k_+, k_2, k_{deg}$) dependent on temperature, calcium ($Ca^{2+}$), and ATP.
2. Cellular: Hodgkin-Huxley neuronal dynamics coupled with microenvironment variables ($Ca^{2+}$, ATP) and amyloid toxicity ($I_{A\beta}$).
3. Network: Tononi spectral complexity ($SC$) calculated via a perturbational complexity index approach.

To overcome hardware limitations, the author introduces three equivalent approximations validated against the complete reference model:

• Izhikevich Surrogate: Replaces the 3-gating-variable HH model with a 2D Izhikevich model, calibrated to match HH dynamics ($R^2 = 0.998$) with a peak error of < 0.12%.
• Voxel Mean-Field: Aggregates $N/V$ neurons (860,000:1 ratio) into voxel states, reducing the state vector size significantly.
• Chunk Streaming: Manages memory by processing the 86B neuron population in 2M-neuron chunks (0.016 GB per chunk), avoiding the need for a full 3.44 TB state vector.
• Probe-Based Complexity: Calculates spectral complexity using a sample of 10,000 neurons rather than the full population, avoiding the formation of an $86B \times 86B$ covariance matrix.

The framework utilizes a multirate integration scheme: fast Izhikevich dynamics ($dt_f = 0.25$ ms) feed firing rates ($\bar{\nu}$) to a slow solver ($dt_s = 100$ ms) that updates amyloid, calcium, and ATP states. An explicit observation map ($R_{PET}$) is defined to separate simulated soluble oligomers ($F2$) from clinical PET plaque readouts (insoluble fibrils, $F$), acknowledging they are distinct observables.

Key Contributions

1. Formalization: Complete ODEs and Bayesian Maximum A Posteriori (MAP) parameters derived from 15 studies without ad hoc tuning.
2. Efficiency: Achieves an 86B-neuron equivalent workload on an 8 GB consumer GPU (RTX 4060) via chunk streaming and mean-field approximations.
3. Speedup: Reduces memory requirements from 3.44 TB to 0.016 GB and wall-clock time for 0.5s biological time from hours to 2.7 seconds (a ~25,400× speedup compared to the complete HH model).
4. Observation Map: Explicitly distinguishes between simulated soluble oligomer reduction and clinical PET plaque burden, preventing direct conflation of model outputs with trial endpoints.
5. Hypothesis Generation: Proposes a "Usnica-Safe" multimechanistic hypothesis involving simultaneous modulation of primary aggregation ($k_n \downarrow$), calcium-dependent aggregation ($Ca^{2+} \downarrow$), and ATP-supported clearance ($ATP \uparrow$).
6. Neonatal Sensitivity: Projects how neonatal $Ca^{2+}$/ATP sensitivity scenarios influence the threshold-crossing time for Alzheimer's pathology over an 80-year lifespan.
7. Open Science: Provides code for fast, slow, and complete models in a public repository.

Results

• Accuracy: The fast model reproduces the reference lag time ($18.32 \pm 1.55$ h vs $18.45 \pm 1.35$ h; MAPE 3.21%) and peak voltage ($115.10$ vs $115.24$ mV; 0.12% error).
• Compound Screening: In a 10-minute GPU run, the framework evaluated 10 compounds. The "Usnica-Safe" candidate showed a simulated 40.9% reduction in soluble oligomers ($F2$) at 50 years, compared to 19.9% for Donanemab and 11.2% for Aducanumab in the model.
• Validation: The model's PET plaque readout predictions for Donanemab, Lecanemab, and Aducanumab matched clinical trial data with map errors between 7.7% and 12%.
• Neonatal Projections: Simulations suggest that high-risk neonatal scenarios (e.g., preterm birth, pollution exposure) could shift the threshold-crossing time for pathology by approximately 20 years over an 80-year lifespan.

Significance and Claims
The paper positions Usnica 4.1 GPU as a tool for democratizing supercomputing for Alzheimer's research, aligning with UN Sustainable Development Goals (ODS 3 and 10). The author explicitly states that the framework enables exploratory screening and hypothesis generation rather than providing evidence of clinical efficacy or safety.

Crucially, the paper emphasizes that:

• The "Usnica-Safe" hypothesis is a model projection, not evidence of long-term clinical safety or efficacy.
• The simulated $F2$ reduction is distinct from clinical PET plaque reduction; the observation map is a calibration tool, not a guarantee of clinical translation.
• The proposed multimechanistic intervention (combining nutraceuticals like EGCG, curcumin, Mg L-threonate, PQQ, and NAD+ support) is a theoretical construct requiring prospective biochemical, toxicological, and clinical validation.
• The model does not account for pharmacokinetics, antibody target engagement, adverse events, regional pathology, tau, neuroinflammation, or cognitive outcomes.

The work fills a gap in the literature by providing the first framework to couple amyloid-HH-spectral complexity dynamics on 86 billion neurons using consumer hardware, while maintaining rigorous mathematical formalization and Bayesian calibration.