Reproduce “Unravelling Causal Genetic Biomarkers of Alzheimer's Disease via Neuron to Gene-token Backtracking in Neural Architecture: A Groundbreaking Reverse-Gene-Finder Approach” on OpenResearch

About this paper

Victor OK Li, Yang Han, Jacqueline CK Lam

The University of Hong Kong researchers introduced "Reverse-Gene-Finder," a methodology that integrates a fine-tuned genomic foundation model with causal tracing and backtracking to identify *causal* genetic biomarkers for Alzheimer's Disease (AD). This approach fine-tuned a Geneformer model for AD classification (AUC 74.67%) and discovered ten novel Most Causal Genes (MCGs), such as PLXDC2 and MT-CO3, validated by literature and pathway enrichment analysis revealing multi-faceted disease mechanisms.

Key insights

• Causal tracing applied to genomic foundation models can effectively localize neurons (MCNs) that mediate causal effects on disease prediction, offering a pathway to model interpretability.
• A novel backtracking mechanism can reverse-engineer the information flow from causally important neurons to identify specific input gene tokens, uncovering previously unacknowledged causal genetic biomarkers.
• The identified MCNs are predominantly located in the front and middle layers of the neural network, suggesting the model's recognition of gene importance involves complex genetic interactions beyond simple expression levels.

Summary by alphaXiv.