Coherence‑Validated Causal World Models for Multi‑Scale Alzheimer’s Disease Progression and Pharmacologic Reversal

David Scott Lewis, Enrique Zueco

Keywords: coherence auditing, Monte Carlo wavelet coherence (MCWC), time–frequency coupling, action-conditioned world models, hierarchical latent dynamics, structured state-space models, counterfactual rollouts, simulator validation, OOD generalization, uncertainty calibration, pessimistic ensembles, mechanistic accountability, interpretable module summaries, agentic causal scaffold, structural causal models (SCMs), differentiable DAG learning, NOTEARS, causal representation learning, regime-shift evaluation, treatment-effect prediction, active experiment selection, Bayesian optimal experimental design (BOED), Alzheimer’s disease dynamics, pharmacologic reversal, NAD+ homeostasis, P7C3-A20, oxidative stress, neuroinflammation, blood–brain barrier integrity, synaptic plasticity

TL;DR: C3WM learns hierarchical action-conditioned dynamics for AD progression and NAD+ reversal, adds an auditable causal scaffold, and uses Monte Carlo wavelet coherence to detect broken couplings and improve OOD counterfactual calibration.

Abstract: Alzheimer’s disease (AD) is a multi-scale dynamical disorder: molecular disruptions in energy and redox homeostasis propagate through cellular stress programs, neuroinflammation, neurovascular dysfunction, synaptic failure, and cognitive decline. Recent interventional evidence that restoring physiological brain NAD+ homeostasis with P7C3-A20 can reverse advanced pathology and behavioral deficits in symptomatic amyloid- and tau-driven mouse models creates an opportunity for action-conditioned simulators that support counterfactual regimen design and mechanistic hypothesis testing. We introduce C3WM (Coherence-Validated Causal World Models), which couples (i) a hierarchical, action-conditioned latent dynamics model spanning molecular, cellular, and tissue/functional scales; (ii) an agentic causal scaffold over interpretable module summaries to support auditable interventional queries; and (iii) a wavelet-coherence auditing layer that evaluates learned simulators as interaction models, not merely forecasters. Our key technical move is to generalize a Monte Carlo Wavelet Coherence (MCWC) validation protocol—initially developed for ground-truth-free causal graph checking—to trajectory-level auditing: we test whether world-model rollouts preserve time–frequency coupling structure between modules (e.g., NAD restoration coupling to oxidative stress and BBB integrity) under held-out intervention schedules. We incorporate coherence losses as regularizers and combine them with conservative (pessimistic) uncertainty estimation to improve out-of-distribution reliability. We present a reversal-centric benchmark specification aligned with published endpoints (NAD+/NADH, proteomic reversal signatures, oxidative damage, neuroinflammation, BBB integrity proxies, synaptic plasticity, and behavior) and evaluate forecasting, counterfactual treatment-effect prediction, regime-shift generalization, mechanistic query stability, and active experiment selection. Across tasks, coherence auditing localizes “good MSE, bad simulator” failure modes and improves calibration for counterfactual rollouts, yielding a practical path from world-model learning to experimentally grounded mechanistic discovery.

Submission Number: 130