ALZHEIMER'S ASSOCIATION
AAIC '26

Title: Iterative Self-Critique for Uncovering Cryptic Epistasis in Alzheimer's: Multi-Hop Genomic Reasoning via Monte Carlo Tree Search
Authors: David Scott Lewis, Enrique Zueco, Haley Yi
Affiliation: AIXC Bio, Zaragoza, Spain, email: reports@aixcbio.com

KEY TAKEAWAY: The Socratic Self-Refine Framework Uncovers Hidden Genetic Interactions in Alzheimer's Using Iterative Self-Critique & Monte Carlo Tree Search

INTRODUCTION
- Background: Complex genetic architecture, tracing subtle multi-hop reasoning tasks. Trace complex cryptic epistatic interactions across multi-hop literature. Standard Retrieval-Augmented Generation (RAG) systems fail.
- Objective: Implement a Socratic Self-Refine framework for fine-grained evaluation and recursive refinement of genomic reasoning.

METHODS
Multi-Hop Genomic Reasoning via Monte Carlo Tree Search (MCTS)
Step 1: Agent tasked with synthesizing multi-omics literature to identify novel genetic interactions.
Step 2: Decompose initial hypotheses into verifiable sub-questions.
Step 3: Guided MCTS: Simulate interactions, generate internal critiques of logical steps.
Step 4: Iteratively prune flawed reasoning paths based on strict evidence retrieval.
Step 5: Learn to correct assumptions over sequential rounds.

RESULTS
Performance on Multi-Hop Tasks: Socratic Self-Refine vs. Standard RAG (higher score is better)
(Bar chart showing higher scores for Socratic Self-Refine)

Reduction in Logical Inconsistencies (%)
(Bar chart showing a significant reduction)

Tracing Cryptic Epistatic Interactions: APOE, TREM2, and microglial inflammatory regulators
(Network diagram showing interactions between APOE, TREM2, APOM, etc.)
- Agent successfully identified previously overlooked gene-gene interactions.
- Reduced logical inconsistencies by 47% compared to single-pass systems.
- Autonomous convergence on plausible hypotheses validated in silico against independent GWAS datasets.

CONCLUSIONS
- Powerful new paradigm for uncovering cryptic genetic architecture in Alzheimer's disease.
- Decomposing hypotheses and systematic pruning enables discovery of hidden interactions in fragmented literature.
- Accelerates novel therapeutic targets rooted in validated multi-gene interactions.