ALZHEIMER'S ASSOCIATION
AAIC '26

Hierarchical Disease-State Generators for Neurodegenerative Genomics: Simulating Multi-Omic Interventions
David Scott Lewis, Enrique Zueco, Haley Yi
AIXC Bio, Zaragoza, Spain
email: reports@aixcbio.com

KEY TAKEAWAY: Hierarchical AI constrains generative diffusion models to obey known biology, enabling realistic, high-throughput in silico screens for single and combination therapies to reverse Alzheimer's disease states.

INTRODUCTION
Black-Box AI Violates Biological Rules vs. Hierarchy-Constrained AI
Neuroinflammation
Alzheimer's pathology (e.g. plaques) -> Disease-Associated Microglia (DAM) -> Homeostatic
- Targeted therapies addressing toxicities need targeted approaches and constrained neurotoxic models
- Standard models have limitations for predicting pathology when aiming to reverse Alzheimer's disease states

METHODS
Single-Cell Multi-Omic Atlases
RNA, ATAC
Patient disease stage, cell type, donor data

Multimodal Disease-State Generator
Hierarchical Regulatory Priors
Hierarchical Regulatory Priors -> Enhancer -> Transcription Factor -> Target Genes -> Core Gene Regulatory Network (GRN)

Conditional Latent Diffusion Architecture -> in silico Counterfactual Cellular States

Simulated Perturbations
Genetic/CRISPR & Pharmacological -> in silico Counterfactual Cellular States

GENERALIZATION ACROSS MULTIPLE NEURODEGENERATIVE CONDITIONS
Disease-State Modules
Multi-omic Data -> Parkinson's Disease Module -> PD Disease-state Counterfactuals
Multi-omic Data -> ALS Module -> ALS Disease-state Counterfactuals
Multi-omic Data -> Huntington's Disease Module -> HD Disease-state Counterfactuals

ADAPTIVE FRAMEWORK
- Same hierarchical constraints retrained with disease-specific GRNs
- Simulate counterfactuals in Dopaminergic, Motor neurons

Candidate Multi-Gene Perturbations
Conformal Prediction for Uncertainty
Low-Confidence
OR Biologically Impossible

RESULTS
Enforced GRN Constraints
Predicted shifts aligned with biology

Strict Hierarchy Fidelity
Predicted shifts vs Proteomic shifts
Simulated counterfactual trajectories

Microglia State Reversal Pathways
(Graph showing Trajectories over Time, DAM)
Simulated counterfactual trajectories simulating reversal trajectories

Candidate Multi-Gene Perturbations
(Graph showing Candidate Multi-Gene Perturbations vs Baselines)
Their expected capacity to shift DAM towards Homeostasis

CONCLUSIONS
Summary of power for recreating this framework
Deep Generative Diffusion vs Hard Mechanistic Constraints
Mechanism-Grounded Decision Support
Triage and Prioritize High-Value Single & Combination Therapies
Save resources before Expensive in vitro or in vivo Validation