INGENE and MODULE broaden the search for schizophrenia-associated genes with co-expression networks
Background: Expression regulation missed by focusing on variants near a gene
A study suggests that using long-range gene-regulatory information from co-expression networks can expand the set of genes whose expression can be predicted and help identify additional genes associated with schizophrenia. Published in Nature Genetics, the study developed two prediction models, INGENE and MODULE, and assessed their benefits when combined with existing models of local regulation.
Transcriptome-wide association studies (TWAS) examine associations between gene expression estimated from genotype data and disease. An expression quantitative trait locus (eQTL) is a genomic location associated with differences in gene expression levels. Existing TWAS mainly use cis-eQTLs near a gene, while effects of distant trans regulation have not been sufficiently captured. Individual trans effects are relatively weak, and the number of combinations to search is large, making them difficult to detect. The researchers narrowed the candidate space using co-expression networks, which describe relationships among genes with correlated expression patterns.
Key findings: Broader prediction coverage and complementary cis and trans information
The researchers used bulk RNA sequencing and genotype data from six regions of human postmortem brains. Bulk analysis measures expression across the mixture of cells in a tissue. INGENE predicts target-gene expression by combining expression levels of co-expressed partner genes estimated with cis models. MODULE uses candidate trans variants associated with the first principal component, a representative expression pattern of a group of co-expressed genes.
In GTEx data, INGENE could predict expression for up to 1.8 times as many genes as the existing CIS model. This does not mean that prediction accuracy for each individual gene was 1.8 times as high. Across brain regions, mean adjusted coefficients of determination (R²) were 7.1–7.8% for CIS, 7.2–9.4% for EpiXcan, 2.0–3.3% for INGENE, and 2.7–4.9% for MODULE. The trans models captured a broader set of genes, but explained less gene-level expression variation than the cis models.
Combining the two types of information helped compensate for their respective gaps. Across multiple brain regions, the full framework significantly predicted expression for 18,744 genes. This total includes cis-only, trans-only, and combined cis–trans predictions, so it does not mean that combining the models improved accuracy for every gene. In a comparison applying an integrated model built with GTEx data to a separate CMC dataset, adjusted R² was also higher than with a single predictor.
The researchers then applied the prediction framework to genotype data from 102,613 people in 62 cohorts in the third Psychiatric Genomics Consortium dataset (PGC3). An analysis of associations between genetically estimated expression and schizophrenia diagnosis identified 766 independent associated genes at a false discovery rate (FDR) below 0.01 after conditional analysis and multiple-testing correction. Of these, 641 associations had not been reported in previous transcriptome-wide association analyses.
Significance and outlook: Expanding the pool of candidates for follow-up functional studies
The study adds regulatory information that may be missed when analyses focus only on variants near a gene, using it to predict expression and search for disease associations. Newly identified associations do not establish that the genes cause schizophrenia or are confirmed treatment targets. The study also did not test treatment response in patients.
The work was based on moderately sized postmortem bulk-tissue datasets, making it difficult to separate effects by cell type, and the analyzed population was limited to people of European ancestry. Although sex was included as a covariate, sex-specific effects on expression regulation were not analyzed separately. The reproducibility of disease associations was assessed by leaving out PGC3 cohorts one at a time; the findings were not confirmed in data independent of PGC. Larger datasets representing more ancestries, cell-type-specific analyses, and functional experiments remain future priorities.
The methods and results are available in the Nature Genetics paper.
Nature Genetics, Published online: 22 June 2026; doi:10.1038/s41588-026-02646-3INGENE and MODULE enhance gene expression prediction by capturing the effects of trans-eQTLs acting within coexpression networks. Integrating cis and trans predictions improves expression imputation for transcriptome-wide association studies.
Teams studying the genomics of schizophrenia could use co-expression network information to search for additional candidates missed by cis-focused TWAS. For example, they could select candidates with association signals in a particular brain region, reassess their associations in independent cohorts, and design cell-model studies to test the functional effects of changes in expression. This is a potential use for prioritizing follow-up experiments.
The paper also provides public locations for the prediction models, analysis results, and code, offering a starting point for replication studies. Whether altering the expression of a candidate gene improves disease outcomes requires separate validation. The current results do not provide a basis for applying these findings to patient diagnosis, drug selection, or dose adjustment.