Agentic Artificial Intelligence-Driven Explainable Deep Learning for Deciphering Noncoding Pathogenic Mechanisms of Delirium Through Genomic Big Data Integration

Abstract

Delirium is a prevalent neuropsychiatric syndrome affecting up to 50% of hospitalized elderly patients, associated with increased mortality, cognitive decline, and health care costs exceeding $164 billion annually. Despite its clinical burden, the genetic architecture underlying susceptibility to delirium remains poorly characterized. We developed an agentic artificial intelligence pipeline integrating genome-wide association study (GWAS) data from FinnGen Release 12 (6854 cases; 384,461 controls) with causal transcriptome-wide association study (cTWAS) using brain-specific eQTL data from Genotype-Tissue Expression. Spatial transcriptomics analysis (gsMap) was performed on the human dorsolateral prefrontal cortex to characterize layer-specific expression patterns. Deep learning models (Enformer, SpliceTransformer) were applied to predict functional consequences of identified variants. GWAS identified a strongly associated locus on chromosome 19q13.32, with lead variant rs429358 (P = 2.79 × 10⁻⁷⁰) defining the apolipoprotein E (APOE) ε4 allele. cTWAS analysis identified two causal genes: APOE (posterior inclusion probabilities [PIP] = 0.871, Z = −8.44) and ZNF226 (PIP = 0.604, Z = 5.63), exhibiting opposite directional effects. APOE downregulation increased susceptibility to delirium, while ZNF226 upregulation conferred an elevated risk. Spatial transcriptomics revealed significant GWAS enrichment in Layer 5 (P < 0.001), with distinct layer-specific expression patterns for both genes. Enformer predicted that rs429358 substantially altered chromatin accessibility at promoter regions (H3K4me3 SNP Activity Difference = 0.30). SpliceTransformer identified significant splice site perturbations (maximum Δ = 0.45). This study provides a comprehensive genetic dissection of delirium through agentic artificial intelligence-driven multiomics integration. We identify APOE and ZNF226 as causal genes with distinct mechanisms and spatial expression patterns. These findings establish a molecular framework for understanding delirium pathophysiology and identify potential therapeutic targets.