Multiscale‑Coherent Representations of Alzheimer’s Disease Reversal via Agentic Active Causal Discovery

Confirmation: I have read and agree with the workshop's policy on behalf of myself and my co-authors.

Track: long paper (4–8 pages excluding references)

Keywords: causal representation learning, Alzheimer’s disease reversal, agentic active causal discovery, NAD+ homeostasis, P7C3‑A20, multiscale coherent representations, structural causal models, Monte‑Carlo wavelet coherence, cross‑species multi‑omics integration, Bayesian optimal experimental design, differentiable causal structure learning, mechanistic interpretability, therapeutic node prioritization, perturbation biology, virtual cell roadmapping

TL;DR: Agentic active causal discovery learns interpretable multiscale‑coherent representations of P7C3‑A20 AD reversal, unifying mouse/human multi‑omics and using Monte‑Carlo wavelet coherence to validate mechanisms and prioritize therapeutic nodes.

Abstract:

Alzheimer's disease (AD) is increasingly defined through multimodal measurements (molecular and cellular programs, biomarker trajectories, and clinical phenotypes), yet most representation learning pipelines treat these views as correlated observations rather than linked causal levels. This gap becomes acute for therapeutic reversal: to learn meaningful representations of life, we must represent not only what covaries with disease, but what can be intervened upon to move the system back toward health. Motivated by recent evidence that restoring brain NAD$^+$ homeostasis can reverse advanced AD phenotypes in multiple mouse models and yields conserved multi-omics signatures across mouse and human brain (Chaubey et al., 2026), we propose a framework for causal, multiscale-coherent representations of reversal.

Our approach couples (i) a multiscale encoder that produces aligned latent states across molecular, cellular, tissue, biomarker, and behavioral levels with (ii) an agentic active causal discovery loop that treats candidate structural causal models (SCMs) as first-class objects. The agent maintains an auditable Causal Graph-of-Thoughts (C-GoT) belief state linking hypotheses, evidence, interventions, and decisions, and selects experiments to maximally reduce uncertainty in clinically meaningful queries (e.g., causal effects on biomarker and cognitive recovery). To address a recurring review concern---validation without experiments---we integrate an independently-developed Wavelet Coherence Validation Protocol as a proxy test for whether proposed causal structures exhibit nontrivial multiscale organization beyond degree-preserving and label-shuffled controls. We further describe spec-gated safety and falsification checks that prevent overconfident claims and support reproducible ``decision cards'' per intervention.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Presenter: ~David_Scott_Lewis1

Format: Yes, the presenting author will definitely attend in person because they attending ICLR for other complementary reasons.

Funding: No, the presenting author of this submission does *not* fall under ICLR’s funding aims, or has sufficient alternate funding.

Submission Number: 44