Agentic Artificial Intelligence-Driven Explainable Deep Learning for Deciphering Noncoding Pathogenic Mechanisms of...

Area of Science:

• Neurogenetics
• Computational Biology
• Genomics

Background:

• Delirium is a common yet poorly understood neuropsychiatric syndrome in elderly patients.
• It leads to significant mortality, cognitive decline, and high healthcare costs.
• The genetic basis of delirium susceptibility remains largely uncharacterized.

Purpose of the Study:

• To genetically dissect delirium susceptibility using an integrated artificial intelligence (AI) approach.
• To identify causal genes and their functional mechanisms contributing to delirium risk.
• To explore the spatial and functional impact of genetic variants in the brain.

Main Methods:

• Developed an AI pipeline integrating genome-wide association study (GWAS) and causal transcriptome-wide association study (cTWAS).
• Utilized brain-specific eQTL data and spatial transcriptomics (gsMap) on human dorsolateral prefrontal cortex.
• Applied deep learning models (Enformer, SpliceTransformer) to predict variant functional consequences.

Main Results:

• Identified the apolipoprotein E (APOE) ε4 allele (rs429358) as a major genetic risk factor for delirium.
• Discovered APOE and ZNF226 as causal genes, with APOE downregulation and ZNF226 upregulation affecting delirium risk.
• Spatial transcriptomics showed significant GWAS enrichment in Layer 5, with distinct gene expression patterns.
• AI models predicted rs429358 alters chromatin accessibility and splice sites.

Conclusions:

• This study provides a comprehensive genetic dissection of delirium using AI-driven multiomics.
• APOE and ZNF226 are identified as key causal genes with distinct molecular and spatial mechanisms.
• Findings establish a molecular framework for delirium pathophysiology and suggest potential therapeutic targets.