From Mouse Reversal to Human Therapeutic Nodes: Verified Causal Reasoning for Post-AGI Biomedicine

Track: Track 1: Technical Foundations for a Post-AGI World

Keywords: Alzheimer's disease, verified causal reasoning, structural causal models (SCM), causal graph-of-thought, Active Causal Hypothesis Testing (ACHT), expected information gain, mouse-to-human translation, multi-omics, therapeutic nodes, verification gates, Wavelet Coherence Validation Protocol (WCVP), wavelet coherence, iPSC-derived neurons and organoids

TL;DR: Mouse-to-human Alzheimer's translation is brittle. Proposes verified causal reasoning: explicit SCMs, Active Causal Hypothesis Testing (info gain), plus verification gates incl. wavelet-coherence checks to stop narrative-only target picks.

Abstract: Alzheimer’s disease (AD) is a high-stakes discovery problem: mechanistic hypotheses are abundant, feedback loops are slow and expensive, and translation from animal reversal to human targets is brittle. A recent study reports pharmacologic reversal of advanced AD phenotypes in transgenic mice and uses multi-omic analysis to propose therapeutic nodes in human brain tissue. We treat this mouse-to-human gap as a causality and verification problem: can a discovery agent synthesize mechanistic and omic evidence into interventions whose assumptions are explicit and whose priorities are justified by falsifiable, quantitative checks rather than narrative plausibility? We present a verified causal reasoning pipeline that couples (i) explicit structural causal models (SCMs) organized via causal graph-of-thought decomposition, (ii) Active Causal Hypothesis Testing (ACHT) to select targets and drugs by expected information gain, and (iii) an “architectural immune system” of verification gates that detect synthetic fallacies, quantify uncertainty, and enforce cross-modal consistency. A central gate is the Wavelet Coherence Validation Protocol (WCVP), an emergent diagnostic in which an agent converts causal-graph traversals into sequences and evaluates multi-scale structural regularity through wavelet coherence against degree-preserving and order-shuffled controls. We outline an evaluation suite spanning synthetic causal benchmarks, STRING-scale graph discovery, and prospective validation in human-relevant perturbation models (iPSC-derived neurons, organoids, or ex vivo slices). Our thesis is that post-AGI biomedical discovery should be judged by verifiability under intervention, and that AD reversal provides an urgent, measurable testbed for such systems.

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: 28