CellSwarm-AD: A Multi-Scale Agent-Based Framework for Virtual Alzheimer's Disease Trials

Computational models of Alzheimer's disease (AD) rarely connect cellular heterogeneity, spatial tissue organization, pathology cascades, and pharmacological intervention within one auditable workflow. We present CellSwarm-AD, a four-layer framework comprising five cell agent classes, a spatial Aβ diffusion environment, an Aβ-Ca²⁺-tau-NF-κB-viability cascade with repository PK/PD models, and an optional experiment orchestration interface. Layer 3 was demonstrated with reproducible prompt templates and deterministic mock outputs; no live large language model was used to generate or modify the quantitative simulation outputs or statistical results. In a prespecified 78-week virtual trial (n = 200 per arm), patient-level repeated measurements were analyzed with Gaussian generalized estimating equations. Week-78 mean (SEM) MMSE-like changes were -1.747 (0.071) for the placebo, -1.368 (0.076) for lecanemab, -1.428 (0.067) for donepezil, and -1.369 (0.065) for independently simulated donepezil plus memantine. The corresponding single-trial Cohen's d values versus the placebo were 0.395, 0.354, and 0.426. Across 20 independent n = 200-per-arm trials, the mean d values were 0.337, 0.359, and 0.326, respectively; replicates were not pooled. Ablation removed most of the treatment contrast when PK/PD was disabled, and fixed-domain grid testing showed decreasing relative L2 error from 6.20% (100 × 100) to 1.49% (200 × 200) against a 400 × 400 reference. These results establish a reproducible proof-of-concept while identifying calibration dependence, weak cross-layer coupling, and the absence of individual-level external validation as current limitations.