Beyond Local Regulation: Network-Based Prediction of Gene Expression and Its Application to Neuropsychiatric Traits

Understanding how genetic variation shapes brain-related traits with complex heritability requires models capable of capturing the multifaceted regulatory architecture of gene expression. Non-coding variants are thought to influence disease risk primarily by modulating transcriptomic regulation rather than through direct coding changes. Gene-level prediction models, which impute expression from genetic data, offer a framework for linking genetic variation to transcriptional consequences; however, current approaches predominantly rely on cis-eQTLs—capturing primarily local regulatory effects—thus limiting their ability to model the distributed architecture of gene regulatory networks. To address these limitations, the present work develops an integrative framework that incorporates gene co-expression network information into genetically regulated expression (GReX) models. By leveraging the structure of transcriptional networks, the approach guides the detection of trans-regulatory (distal) effects and improves transcriptomic prediction. In the first study, I implemented this framework through the development of two complementary algorithms, INGENE and MODULE, which integrate co-expression module structure into trans-eQTL selection and dimensionality reduction for gene-level prediction. Using RNA-seq and genotype data from postmortem brain cohorts (LIBD, CMC, and GTEx), I trained and validated these models across six brain regions (amygdala, caudate nucleus, dorsal/subgenual anterior cingulate cortex, dorsolateral prefrontal cortex, hippocampus). Benchmarking against both an original cis-based model and EpiXcan—the leading benchmark for cis-model performance on our training dataset—demonstrated that the integration of cis- and trans-predictions significantly improves gene coverage and predictive accuracy across independent datasets. In the second study, I applied the co-expression-informed prediction models to large-scale schizophrenia (SCZ) cohorts from the Psychiatric Genomics Consortium wave 3 (PGC3) to evaluate their utility in gene-trait association discovery. By imputing gene expression across brain regions and performing association testing, I identified 1,764 SCZ-associated genes across regions (pFDR < .01), including 1,515 novel associations not captured by cis-only approaches.