Leveraging machine learning and network biology approaches to predict brain gene expression from blood transcriptomes

BACKGROUND: Blood-based biomarkers offer a promising non-invasive strategy for detecting disease-related changes and monitoring tissue and organ health, including brain function. While recent studies have leveraged blood transcriptomic data to predict gene expression in the brain, existing models generally suffer from poor accuracy, limiting their translational utility. FINDINGS: We present an integrative prediction system (IPS) that combines machine learning with network biology to predict region-specific brain gene expression from blood transcriptomic data. Our framework integrates global blood transcriptomic signals, co-expression network features, and inter-tissue gene-gene interaction data linking blood genes to their target genes in the brain. Applied to the Genotype-Tissue Expression cohort, IPS substantially outperforms existing approaches in both the number and accuracy of brain genes that can be reliably predicted from blood. Notably, immune-related blood genes emerged as key contributors to model performance, underscoring the systematic interplay between peripheral immune signaling and central nervous system. CONCLUSIONS: These findings highlight the potential of blood-based transcriptomic models as scalable, non-invasive tools for studying brain function and developing diagnostic and prognostic biomarkers for neurological and psychiatric disorders.