Co-expression-based models improve eQTL predictions for transcriptome-wide association studies and highlight new schizophrenia-associated genes Authors: Fabiana Rossi, Leonardo Sportelli, Gianluca C Kikidis, Giulia Grassi, Fabio Di Camillo, Alessandro Bertolino, Giuseppe Blasi, Christopher J Borcuk, Daniela Fusco, Thomas M Hyde, Joel E Kleinman, Davide Marnetto, Silvia Pellegrini, Antonio Rampino, Benedetto Vitiello, Stephan Ripke, Alice Braun, Julia Kraft, Sintia Iole Belangero, Paulo R Menezes, Celso Arango, James T R Walters, Michael C O’Donovan, Michael J Owen, David Braff, Aiden Corvin, Derek W Morris, Enrico Domenici, Jim van Os, Esref Atbaşoğlu, Meram C Saka, Marta Di Forti, Bernhard T Baune, Carlos N Pato, Andrew McQuillin, Vera Golimbet, Nikolay Kondratyev, Valentina Escott-Price, Anna Gareeva, Elza Khusnutdinova Published: 2026 ABSTRACT AbstractMost genetic variants associated with complex heritability phenotypes lie in non-coding regions and are thought to influence disease risk by regulating gene expression. However, most transcriptome-wide association approaches primarily model local (cis) genetic effects, leaving much of gene regulation unexplained. Here, we show that incorporating distal (trans) regulatory effects improves the prediction of gene expression and the identification of disease-associated genes. Using RNA sequencing data from six human post-mortem brain regions, we developed INGENE and MODULE, two models capturing the combined influence of candidate trans-acting variants within gene coexpression networks. Integrating these models with conventional cis-based predictors improved gene expression imputation (maximum likelihood estimation, α = 0.05) for 18,744 genes across regions. Applying this framework to Psychiatric Genomics Consortium wave 3 genotypes identified 766 genes associated with schizophrenia (PFDR < 0.01), including 641 not previously reported by transcriptome-wide analyses. These findings highlight the contribution of distal regulatory mechanisms and gene network interactions to schizophrenia risk.Subject terms: Schizophrenia, Gene expression profiling Main Schizophrenia (SCZ) is a severe psychiatric disorder with twin-based heritability estimates between 60% and 80%1,2. Genome-wide association studies (GWAS) have identified numerous common single nucleotide polymorphisms (SNPs) associated with SCZ risk. Although increasing GWAS sample sizes have enhanced the number of significant loci3,4, the explained heritability remains limited, and translating these statistical associations into molecular mechanisms remains challenging5. Most risk variants are non-coding2 and are probably associated with gene expression6,7. Gene expression provides a functional intermediary between the DNA sequence and disease liability6,8. Accordingly, transcriptome-wide association studies (TWAS)9 integrate GWAS with expression quantitative trait loci (eQTL) reference panels to impute the genetic component of gene expression and test its association with phenotypes9–14. In TWAS, two main approaches are commonly used: the direct prediction of expression in genotyped samples (gene-level TWAS) and the indirect estimation of associations between predicted expression and traits, accounting for linkage disequilibrium among SNPs (SNP-level TWAS). Both approaches identify gene-level trait associations while reducing the multiple-testing burden compared to SNP-based GWAS11,12. However, most existing TWAS implementations, including PrediXcan11 and related methods12,15,16, mainly rely on cis-eQTLs (variants within ±1 Mb of a gene). This assumption limits interpretability, as most non-coding risk variants do not colocalize with known eQTLs6,17–19, and cis effects explain only a fraction of expression heritability2,20–26. Trans regulation may be particularly relevant in SCZ, whereby risk variants act through coordinated transcriptional programs across brain regions and developmental stages27,28. As trans effects are weaker than cis effects, their detection requires larger sample sizes11, and systematic characterization remains challenging17. Nonetheless, risk-associated variants are thought to converge on coordinated transcriptional programs28 across brain regions, over time29,30 and even at the cell-type level31–33. Several studies have reported that SCZ-associated genes converge within coexpression networks33–39, suggesting that shared variance among sets of genes may reflect underlying regulatory architecture and disease-relevant biology, and these findings can be validated in vitro39,40. Several studies have leveraged SNPs proximal to genes in coexpressed sets, termed coexpression-based eQTLs (co-eQTLs), to generate system-level polygenic scores to translate coexpression into phenotypic predictions15,38,41–47. These approaches leverage gene modules or coexpression partners to prioritize variants that may exert coordinated regulatory effects39,48,49. Given that it limits the search space to gene sets, this approach to identifying trans-eQTLs differs from genome-wide scans, which are often underpowered owing to the burden of multiple comparisons11. Here, we developed two predictive frameworks to quantify trans-regulatory contributions to gene-level expression: the imputed network gene-expression trans-eQTL (INGENE), which models the trans effects of cis-regulated coexpression partners on a target gene; and the module quantitative trait loci eigengene (MODULE), which identifies SNPs associated with the eigenvalue (first principal component) of a gene’s coexpression module. In addition, we implemented two cis-based benchmarks: a standard elastic-net model11 (CIS) and EpiXcan12. We evaluated whether INGENE and MODULE increase the number of genetically predictable risk genes beyond cis-based models and whether integrating cis and trans components enhances predictive accuracy. We then applied our framework in a coexpression-informed TWAS (coTWAS) of SCZ using 62 Psychiatric Genomic Consortium (PGC) wave 3 (ref. 5) cohorts. The overall study design is summarized in Fig. 1. Analytical framework for cis-informed and trans-informed gene expression prediction and trait association. The workflow consists of four stages. (1) Model training: cis-based (CIS, EpiXcan) and trans-based (INGENE, MODULE) elastic-net models were trained using RNA-seq and genotype data from six LIBD post-mortem brain regions. CIS and EpiXcan model local (±1 Mb) eQTL effects, whereas INGENE leverages cis-regulated coexpression partners and MODULE captures SNPs associated with module eigengenes (co-eQTLs). (2) Model testing: trained weights were applied to independent genotype datasets (GTEx and CMC) to impute gene-level expression and evaluate cross-cohort reproducibility. (3) Score integration: cis-based and trans-based predictions were combined to generate unified gene-level expression estimates. (4) Gene–trait association: integrated genetically imputed expression was tested for association with SCZ diagnosis across 62 PGC3 cohorts through coTWAS. CN, caudate nucleus; ACC, anterior cingulate cortex; HP, hippocampus. Results INGENE and MODULE increase the number of predicted genes across brain regions Evaluation of cis and trans model training performance We trained models using the Lieber Institute for Brain Development (LIBD) brain RNA sequencing (RNA-seq) dataset, which provides the largest available sample sizes across the examined datasets (Table 1). As benchmarks, we implemented a standard elastic-net model11 (CIS) and EpiXcan12, which incorporates Roadmap Epigenomics50 chromatin priors (except for the amygdala). Across five regions, EpiXcan increased both the number of predictable genes and training accuracy relative to CIS (mean adjusted R2 of 0.153 vs 0.140; ~9% relative gain; Fig. 2a and Supplementary Fig. 1). However, the models showed partial complementarity: approximately 6–8% of genes were uniquely captured by either approach, or CIS outperformed EpiXcan for a subset of genes in each region (caudate nucleus, 1,558; dorsal anterior cingulate cortex (dACC), 1,569; dorsolateral prefrontal cortex (DLPFC), 1,562; hippocampus, 1,736; subgenual anterior cingulate cortex (sACC), 1,193; Fig. 2b). Both cis models were therefore retained in downstream analyses. Table 1Sample characteristics of brain transcriptomic datasets used for model training and validationBrain regionLIBDGTExCMCnAge (mean ± s.d.)Sex (% female)AncestrynAge (mean ± s.d.)Sex (% female)AncestrynAge (mean ± s.d.)Sex (% female)AncestryACCdACC17647 ± 1535.3EUR14160 ± 1028.4EUR15961 ± 1833%EURsACC50847 ± 1631EURAmygdala46147 ± 1629EUR11458 ± 1030%EURCN21150 ± 16.425.1EUR19959 ± 1025%EURDLPFC58446 ± 1531.5EUR16059.3 ± 927%EUR38770 ± 1740%EURHP23646 ± 1623EUR14559 ± 1126.2%EURThe table summarizes sample size and demographic characteristics of post-mortem brain tissue datasets from the LIBD, GTEx and CMC across brain regions. For each dataset and region, the number of samples (n), mean age at death (in years), percentage of female donors and ancestry are reported. EUR, European; CN, caudate nucleus; HP, hippocampus. Sample characteristics of brain transcriptomic datasets used for model training and validation The table summarizes sample size and demographic characteristics of post-mortem brain tissue datasets from the LIBD, GTEx and CMC across brain regions. For each dataset and region, the number of samples (n), mean age at death (in years), percentage of female donors and ancestry are reported. EUR, European; CN, caudate nucleus; HP, hippocampus. Comparative performance of cis and trans prediction models across brain regions. a, Number of predictable genes for CIS and EpiXcan across LIBD brain regions. b, Gene-level comparison of cross-validated adjusted R2 between CIS and EpiXcan for the top 20 genes per brain region ranked by improvement in cross-validated adjusted R2. Each point represents one gene (n = 20 per region), and paired points are connected by lines. c, Number of predictable genes across brain regions for all models: CIS, EpiXcan, INGENE and MODULE. Trans-aware models substantially expand the number of genes whose expression can be genetically predicted. d, Overlap of predicted genes across models. Left, intersection among CIS, EpiXcan, INGENE and MODULE. Right, overlap of trans-aware models with the union of cis predictors (CIS + EpiXcan). We next compared the cis-based predictors (CIS and EpiXcan) with the trans-informed frameworks, INGENE and MODULE. Trans models were trained in all available post-mortem datasets, including LIBD, the Genotype-Tissue Expression project (GTEx) and the CommonMind Consortium37 (CMC) (Table 1). Only genes replicated across datasets were retained (Supplementary Notes and Supplementary Fig. 2). This cross-dataset filtering reduced false positives and improved stability of trans-eQTL signals. Cis models were not penalized by this filtering, as we observed that performance was sensitive to training sample size (Supplementary Notes and Supplementary Figs. 3 and 4). We next quantified training performance across models, revealing heterogeneous predictive capacity across brain regions (Fig. 2c). When pooling regions, CIS predicted 26,285 genes, EpiXcan predicted 25,702, INGENE predicted 18,403 and MODULE predicted 16,721 (Fig. 2d left). Cis-based models contributed 8,797 unique predictions (30% of total), whereas INGENE and MODULE showed substantial overlap, sharing 16,103 common predictions (96% of MODULE; 88% of INGENE; Fig. 2d right). Supplementary Fig. 5 shows the unfiltered performance comparison. To examine whether trans-predicted genes shared regulatory architecture across frameworks, we assessed whether SNPs acting as trans-eQTLs for a MODULE-predicted gene also serve as cis-eQTLs for coexpression partners used by INGENE to predict that same target gene. Across brain regions, 15–29% of MODULE-predicted genes exhibited such cis–trans regulatory relationships (Supplementary Table 1), indicating partial convergence between cis-mediated effects and module-level regulatory structure. Cross-validated adjusted R2 comparisons for 15,232 commonly predicted genes across models (Fig. 2d right and Supplementary Fig. 6) showed that EpiXcan outperformed CIS (mean delta (Δ)R2 = 0.0103, one-tailed Wilcoxon signed-rank test, P = 2.06 × 10−20) and INGENE achieved higher mean adjusted R2 than CIS (ΔR2 = 0.0241) and EpiXcan (ΔR2 = 0.0257), although these differences were not statistically significant. By contrast, MODULE significantly outperformed all other models, with improvements relative to INGENE (ΔR2 = 0.0464, P = 5.04 × 10−16), CIS (ΔR2 = 0.0721, P = 3.02 × 10−3) and EpiXcan (ΔR2 = 0.0705, P = 1.46 × 10−4). To contextualize our approach, we compared it with published cis–trans frameworks (Bayesian genome-wide (BGW)-TWAS51 and MOSTWAS52; Supplementary Notes), restricting analyses to the DLPFC, the only common region. A gene was classified as cis or trans when the corresponding component contributed to its prediction (Supplementary Fig. 7a). Approximately 59% of BGW cis–trans predictions and 96% of the cis–trans set in our framework were recovered by both approaches (Supplementary Fig. 7b), indicating strong concordance in regulatory classification while extending predictions to additional genes. Evaluation of cis and trans models on a testing dataset We assessed the replication of all four models—CIS, EpiXcan, INGENE and MODULE—in the independent GTEx cohort across the same brain regions used for model training. We applied each model to GTEx genotype data to generate predictions, thereby avoiding potential overfitting by evaluating performance in an independent testing dataset. We evaluated predictions based on a positive correlation between predicted and observed expression (Pearson’s r > 0) and adjusted R2 > 0. In GTEx, CIS predicted between 4,732 and 7,777 genes across regions. EpiXcan predicted 4,858–7,797 genes. INGENE achieved broader coverage, predicting 5,429–10,749 genes, while MODULE predicted 5,084–10,718 genes (Fig. 3a). Therefore, both trans models substantially increased the number of predictable genes relative to cis-only approaches, with INGENE predicting up to 1.8-fold more genes than CIS and 1.7-fold more than EpiXcan (Supplementary Fig. 8). Predictive models replicate across brain regions in the external GTEx dataset and predict sets of genes with different performance. a, Barplots showing the number of predicted genes in the GTEx dataset by CIS, EpiXcan, INGENE and MODULE models. b, Boxplots of adjusted R2 values in predicting gene-level expression in GTEx using CIS, EpiXcan, INGENE and MODULE predictive models across brain regions. The median is represented by the central line, with the interquartile range (IQR) as the box. Whiskers extend to 1.5× the IQR, and outliers are plotted as individual points. Each point represents a gene whose expression was predicted in the independent GTEx dataset. The number of genes included per model and region is as follows: amygdala: CIS (n = 5,936), INGENE (n = 8,052), MODULE (n = 10,057); CN: CIS (n = 6,140), EpiXcan (n = 6,172), INGENE (n = 6,228), MODULE (n = 6,016); dACC: CIS (n = 5,284), EpiXcan (n = 5,447), INGENE (n = 9,537), MODULE (n = 8,455); DLPFC: CIS (n = 7,777), EpiXcan (n = 7,797), INGENE (n = 9,464), MODULE (n = 8,895); HP: CIS (n = 4,732), EpiXcan (n = 4,858), INGENE (n = 5,429), MODULE (n = 5,048); sACC: CIS (n = 6,782), EpiXcan (n = 6,799), INGENE (n = 10,749), MODULE (n = 10,718). c, Venn diagram showing the number of total intersecting predicted genes among models. d, Barplots showing the percentage corresponding to ‘n’ genes predicted within each region for which CIS or EpiXcan achieves the higher adjusted R2 among genes predicted by both models. As expected, cis-based models explained a higher proportion of expression variance per gene than trans-informed models. In GTEx, mean adjusted R2 ranged from 7.1% to 7.8% for CIS and from 7.2% to 9.4% for EpiXcan, aligning with previous findings12. By contrast, INGENE explained between 2% and 3.3% of variance, and MODULE explained between 2.7% and 4.9% (Fig. 3b and Supplementary Fig. 9). Across regions, EpiXcan significantly exceeded INGENE and MODULE in mean adjusted R2 (one-tailed Wilcoxon test, all P < 2.2 × 10−16; Supplementary Fig. 10). Compared to CIS, EpiXcan showed significantly higher performance in caudate nucleus, dACC, DLPFC and hippocampus (all one-tailed Wilcoxon tests, P < 2.2 × 10−16), whereas no significant difference was observed in sACC (P = 0.64). Notably, testing performance differed from training estimates, highlighting the importance of external validation. Despite the substantial overlap across frameworks (Venn diagram in Fig. 3c), each method contributed unique predictions in GTEx: CIS identified 561 genes not captured by other models, EpiXcan identified 289, INGENE identified 299 and MODULE identified 244. This complementarity supports retaining all four models within the integrative framework. For genes predicted by both CIS and EpiXcan, we selected the model with the higher adjusted R2 (Fig. 3d) for downstream analyses, maximizing predictive accuracy while avoiding redundancy. To further assess reproducibility, we applied all models to the CMC post-mortem datasets (DLPFC and ACC; Table 1) and quantified cross-cohort consistency relative to GTEx (Supplementary Fig. 11). CIS retained 2,136–4,535 common genes (32–47% overlap) with cross-cohort correlations of r = 0.43–0.68, and EpiXcan showed similar replicability (2,183–4,565 genes, 33–47%; r = 0.43–0.68). INGENE yielded 11,057–12,274 shared genes (77–79%) with r = 0.17–0.52, whereas MODULE retained 17,594–19,385 genes (~88%) with r = 0.14–0.17 (Supplementary Fig. 11). Although cis predictors exhibited higher correlation, trans-informed models demonstrated a higher percentage of shared genes across independent cohorts. We further benchmarked our LIBD-trained models against BGW51 and MOSTWAS52 in DLPFC across both CMC and GTEx datasets (Supplementary Notes). Across cohorts, our models predicted the largest number of genes and showed comparable or higher adjusted R2 distributions (Supplementary Fig. 7c,d). The broader performance range observed for BGW-TWAS and MOSTWAS probably reflects their integration of cis and trans components, whereas INGENE and MODULE, which rely exclusively on trans signals, showed narrower and more stable distributions. Notably, the median predictive accuracy of INGENE and MODULE exceeded that of the other trans-informed models. Taken together, these results show that our coexpression-based approach improves gene coverage while maintaining replicability across independent datasets. Functional genetics of co-eQTLs To define the regulatory architecture of trans-associated variants, we examined MODULE-derived trans-SNPs that were also significant cis-eQTLs in the GTEx multi-tissue QTL resource53 (Supplementary Table 3). Depending on the region, 25–44% of trans-SNPs overlapped with GTEx cis-eQTLs and were associated with 5,821–19,276 cis-regulated eGenes (Supplementary Table 3). These results show that a substantial portion of the trans signals detected by MODULE arises from SNPs that also exert cis effects on nearby genes. Gene Ontology enrichment of these cis-eGenes revealed overrepresentation of ATP-dependent and catalytic activities, electron-transferase functions and protein-binding categories, including major histocompatibility complex (MHC) class II binding and cadherin binding (false discovery rate (FDR) < 0.05; Supplementary Fig. 12). Regulomic analysis identified 252 transcription factors significantly enriched among GTEx cis-eGenes linked to MODULE trans-SNPs (Bonferroni P < 0.05; Supplementary Fig. 13a). Enriched transcription factors included GATAD2A, RERE, IRF3 and SP4, genes previously prioritized in SCZ GWAS5 (Supplementary Fig. 13b). Together, these findings suggest that a subset of trans-eQTLs reflects coordinated regulatory programs involving transcriptional control mechanisms. Functional enrichment of SCZ risk variant-associated genes in predictive models We next assessed whether regulatory SNP weights were enriched among SCZ-associated variants5 with model-derived SNP weights (defined as the mean absolute weight across all regulated genes). We did not perform the same analysis for INGENE, as its predictions derive from cis-regulatory SNPs of coexpressed partner genes rather than direct SNP-to-gene associations, and therefore reflect propagated partner-gene cis effects rather than regulation of the target gene itself. Across brain regions, both cis-based and trans-based models showed significant positive correlations between QTL weights and the effect sizes of SCZ-associated variants (Supplementary Table 4 and Supplementary Fig. 14), indicating that SNPs with stronger regulatory effects are associated with higher disease risk. CIS and EpiXcan showed comparable correlations across regions (r = 0.14–0.18), whereas MODULE consistently achieved markedly stronger correlations (r = 0.28–0.42) (Supplementary Table 4). To assess whether these differences were significant, correlations were compared using Fisher’s z-test, confirming that MODULE outperformed CIS and EpiXcan across regions (P < 2 × 10−16; Supplementary Fig. 15). These results suggest a stronger association of trans-eQTLs than for cis-eQTLs with common variant pathogenicity. Similar trends were observed for bipolar and major depressive disorders (Supplementary Fig. 14). We further tested whether trans-predicted genes that are enriched for SCZ-associated SNPs, quantified by the PGC weight ratio (Supplementary Notes), showed increased connectivity27 to PGC3-prioritized genes5 across regions. When genes were ranked by PGC weight quintiles, mean connectivity increased monotonically, as confirmed by linear regression, Spearman’s correlation and permutation testing (n = 1,000 permutations per model-region pair; Supplementary Fig. 16, Supplementary Table 5 and Supplementary Notes). MODULE showed robust enrichment across DLPFC, caudate nucleus, hippocampus and sACC (all P < 0.001), whereas CIS exhibited a weaker effect in the amygdala (P = 0.037) and EpiXcan showed no significant enrichment. These results indicate that trans-derived signals preferentially localize to coordinated coexpression networks linked to SCZ risk27. Combination of cis and trans scores enhances gene expression prediction in GTEx We next tested whether combining cis and trans predictions improves gene expression imputation in GTEx. For genes with both cis and trans components, we compared two nested models: a baseline model including covariates and cis-derived predictions; and a full model including covariates, cis-predicted expression and trans-predicted expression. Model comparison was performed using maximum likelihood estimation (α = 0.05). Genes were considered enhanced by trans predictions when the inclusion of the trans component led to a significant increase in adjusted R2 relative to the cis-only model. Across brain regions, integration of cis and trans predictors increased the number of significantly predictable genes (Fig. 4a). For these genes, inclusion of the trans term in the maximum likelihood estimation model consistently yielded a significant increase in variance explained compared to the cis-only model. This suggests that the trans component provides additional explanatory power that local genetic regulation alone does not capture (Fig. 4b). When pooling regions, 18,744 genes were significantly predicted by the integration of all models (Fig. 4a). INGENE and MODULE trans scores improve gene prediction across brain regions. a, Number of genes significantly predicted by cis-only, trans-only and cis–trans models across regions; gray indicates unique genes when pooling regions. Best-fitting model counts per region: amygdala (trans, 7,169; cis, 2,717; cis–trans, 847), CN (trans, 5,763; cis, 3,674; cis–trans, 889), dACC (trans, 8,393; cis, 2,822; cis–trans, 949), DLPFC (trans, 8,148; cis, 4,475; cis–trans, 1,362), HP (trans, 5,790; cis, 2,654; cis–trans, 881), sACC (trans, 8,521; cis, 3,443; cis–trans, 1,232). b, Distribution of ΔR2 = adjusted R2 (cis + trans) − adjusted R2 (cis) for genes for which cis–trans was the best model. Violin plots show kernel density. Boxplots show the median (center line), IQR (box bounds, 25th–75th percentiles), whiskers extending to the most extreme values within 1.5× the IQR and points indicating outliers beyond the whiskers. The dashed horizontal line indicates ΔR2 = 0. Median ΔR2: amygdala, 0.0418 (n = 847); CN, 0.0238 (n = 889); DLPFC, 0.0288 (n = 1,362); HP, 0.0311 (n = 881); dACC, 0.0348 (n = 949); sACC, 0.0337 (n = 1,232). c, Distribution of ΔR2 = adjusted R2 (combined) − adjusted R 2 (single model) in the independent CMC dataset, stratified by brain region and model (CIS, EpiXcan, INGENE and MODULE). Violin and boxplots are defined as in b. Each point represents one gene. Statistical significance was assessed using one-sided Wilcoxon signed-rank tests (H1: combined > individual). Exact P values and sample sizes were: DLPFC: CIS, n = 3,875, P = 6.30 × 10−80; EpiXcan, n = 3,935, P = 5.23 × 10−79; INGENE, n = 8,543, P < 2.2 × 10−308; MODULE, n = 7,026, P < 2.2 × 10−308. dACC: CIS, n = 1,602, P = 9.66 × 10−201; EpiXcan, n = 1,633, P = 1.03 × 10−207; INGENE, n = 7,575, P < 2.2 × 10−308; MODULE, n = 6,663, P < 2.2 × 10−308. sACC: CIS, n = 2,108, P = 4.44 × 10−234; EpiXcan, n = 2,108, P = 2.71 × 10−236; INGENE, n = 8,222, P < 2.2 × 10−308; MODULE, n = 7,608, P < 2.2 × 10−308. Having shown that trans components improve prediction for a subset of genes, we next constructed a unified model combining cis-derived and trans-derived predictions. Given that cis-SNPs were excluded when generating trans predictors, these components were independent. For genes with at least two predictors, we fit a linear model in GTEx using observed expression as the outcome and CIS, EpiXcan, INGENE and MODULE predictions as explanatory variables. We evaluated model performance in the CMC (dACC, DLPFC and sACC). Combining cis-based and trans-based predictions into a single model (‘combined’ in Fig. 4c) improved gene expression prediction across all regions, explaining 4.1–5.0% of variance across 7,640–8,600 genes. By comparison, CIS and EpiXcan accounted for 2.3–6.4% of variance in 1,602–3,935 genes, while INGENE and MODULE explained 1.8–2.9% and 1.4–3.7% across 6,663–8,543 genes, respectively. Across regions, the combined model yielded significantly higher adjusted R2 values than any individual predictor (one-sided Wilcoxon signed-rank test; Fig. 4c). We therefore used the combined model for subsequent gene–trait association analyses. coTWAS identifies trait associations in PGC3 cohorts To identify gene associations with SCZ, we applied the integrated cis–trans framework to 62 PGC3 (ref. 5) cohorts (102,613 individuals; Supplementary Table 6). Within each cohort and brain region, we retained genes that met GTEx selection criteria (Supplementary Fig. 17). We associated predicted expression with SCZ diagnosis using logistic regression followed by meta-analysis of effect sizes across cohorts (coTWAS). To identify independent signals, we performed conditional analysis13 across 315 genomic regions (±1 Mb window) with FDR-corrected results (FDR = 0.01) to account for overlapping subjects across brain regions. Across 96,535 tests, we identified 1,162 significant associations (975 non-MHC), corresponding to 766 independent genes (693 non-MHC; Table 2 and Supplementary Table 7). Of these, 381 genes were genetically upregulated and 414 were downregulated in SCZ (Fig. 5). Table 2Summary of significant coTWAS gene-level associations across brain regionsBrain regionTotal number of testsNumber of significant genesβ direction (positive/negative)Total number of unique genesAmygdala96,53515778/79766CN15173/78dACC18298/84DLPFC232130/102HP15771/86sACC283120/163The table reports the number of genes significantly associated with SCZ in each brain region following conditional coTWAS analysis. Total number of tests indicates the total number of gene-level conditional association tests performed across all regions. Number of significant genes denotes the number of genes reaching significance in each region after FDR correction (FDR < 0.01). β direction indicates the number of genes with positive or negative effect sizes (β), corresponding to higher or lower genetically predicted gene expression associated with cases, respectively. Total number of unique genes indicates the total number of distinct genes identified across all regions combined. The unit of analysis is the gene. Summary of significant coTWAS gene-level associations across brain regions The table reports the number of genes significantly associated with SCZ in each brain region following conditional coTWAS analysis. Total number of tests indicates the total number of gene-level conditional association tests performed across all regions. Number of significant genes denotes the number of genes reaching significance in each region after FDR correction (FDR < 0.01). β direction indicates the number of genes with positive or negative effect sizes (β), corresponding to higher or lower genetically predicted gene expression associated with cases, respectively. Total number of unique genes indicates the total number of distinct genes identified across all regions combined. The unit of analysis is the gene. Gene-level conditional association results (coTWAS) across brain regions. Chromosomal position and −log10 conditional P values after FDR correction (FDR < 0.01). Conditional association statistics were derived using the coTWAS framework. Each point represents one gene identified as significant in the conditional analysis (amygdala, n = 157; CN, n = 151; dACC, n = 182; DLPFC, n = 232; HP, n = 157; sACC, n = 283; total, n = 766). The dashed horizontal line indicates the significance threshold. Genes above the dashed line represent positive effects (upregulated in cases), and those below the line represent negative effects (downregulated in cases). For clarity, a maximum of 50 gene names per region are shown. Leave-site-out validation across cohorts identified 556 replicable genes (empirical one-tailed P < 0.05; Supplementary Fig. 18 and Supplementary Table 7) with a mean (±s.d.) concordance rate of 69.4% ± 5.1% and mean cross-site effect-size correlation R2 = 0.34 ± 0.11. These results indicate that about 70% of site-specific effect directions were consistent across leave-out iterations, supporting consistency across cohorts. Gene Ontology enrichment revealed regionally structured biology (Fig. 6a). In the DLPFC, enrichment across β directions converged on synapse organization, particularly AMPA receptor clustering, and membrane or vesicle transport pathways, including clathrin-coated vesicles, late endosome and lysosomal membrane, indicating altered receptor trafficking. Antigen-processing and MHC pathways were significantly enriched and upregulated (β > 0) in DLPFC and dACC and remained significantly enriched across regions when all β directions were considered, supporting their widespread role in individuals with SCZ. These findings delineate regionally stratified biology, with excitatory synaptic remodeling predominating in prefrontal cortex and immune-related processes distributed across cortical and limbic regions. Integrative analysis of coTWAS-identified genes and SCZ-associated gene sets. a, Gene Ontology enrichment analysis of coTWAS-significant genes across brain regions. Enrichment significance was assessed using a one-sided hypergeometric test, with FDR correction applied for multiple comparisons. The x axis shows a −log10 scale of adjusted FDR P values for the top enriched term, and the y axis shows functional categories grouped by Gene Ontology domains. Marker colors indicate the direction of the genetic effect; marker shapes denote ontology categories. b, Overlap between coTWAS-identified genes and previously reported SCZ-associated gene sets. The Venn diagram (top) shows the intersection between coTWAS genes and genes reported in previous studies, including Huckins et al.13, SCHEMA61, PGC3 Mendelian randomization (MR), PGC3 fine-mapping (FM) and PGC3-prioritized genes5, Borcuk et al.27, Gandal et al.58, Gusev et al.62, Hall et al.59 and Collado-Torres et al.60. The matrix (bottom) shows pairwise overlaps quantified by the Szymkiewicz–Simpson overlap coefficient, in which darker blue colors indicate larger proportional overlap values. c, Presence–absence heatmap of genes present in at least four studies (n ≥ 4) across the indicated datasets. Rows represent genes and columns represent studies. Blue squares indicate that the gene was reported in the corresponding study, whereas white squares indicate its absence. d, Fold-enrichment versus −log10(P) for the overlap between coTWAS genes (n = 766) and each SCZ-associated gene set. Point size indicates the number of overlapping genes. Marker color represents −log10(P values), with darker colors indicating stronger statistical significance. Marker shapes distinguish data types: diamonds represent gene sets and circles represent TWAS-derived gene lists. The dashed horizontal line indicates the nominal significance threshold (P = 0.05). Enrichment was assessed using a one-sided hypergeometric test for overrepresentation. Among coTWAS significant genes, 354 were predicted by a single model; CIS accounted for 21%, EpiXcan accounted for 20%, INGENE accounted for 40% and MODULE accounted for 19%. Across regions, cis components represented 26.5–54.3% of associations, cis–trans represented 17.2–30% and trans-only represented 18.5–44.5% (Supplementary Fig. 19). MAGMA z-score (Supplementary Notes) correlated with cis-predicted genes (r = 0.10, P = 0.003) but not with trans-only predictions (Supplementary Fig. 20; r = −0.02, P = 0.78), consistent with MAGMA’s underlying design, which focuses on SNPs located proximal to genes. Complementarity was further illustrated by genes showing strong coTWAS evidence (log10(adjusted P) > 4) but weak MAGMA support (|z| < 4; Supplementary Fig. 21). Cell specificity enrichment using a human atlas54 revealed region-specific patterns (Supplementary Fig. 22) of transcriptional perturbation. In the amygdala, upregulated genes (β > 0) were overrepresented in excitatory DG (exDG)-like neurons. In dACC, neuronal stem cells showed enrichment for upregulated genes, while exDG-like neurons were enriched among downregulated genes (β < 0). In DLPFC, pyramidal neurons and exDG-like neurons exhibited upregulation, whereas GABAergic interneurons (GABA2) showed downregulation. These results suggest region-specific and cell-type-specific transcriptional alteration, consistent with neuronal dysfunction in SCZ. coTWAS identified several previously reported SCZ-related genes, including SNX19 (ref. 55), CYP21A1P56 and C4A57 (downregulated across all regions), and GATAD2A (upregulated in DLPFC). These genes were predominantly driven by cis components, with additional region-specific contributions from trans models. Comparison with prior studies5,13,27,58–62 showed that 125 genes had been previously reported (Fig. 6b), whereas 641 (83.7%) were novel (Supplementary Fig. 23a). Known associations were largely recovered by cis and cis–trans models (28.8–56%), while novel associations were enriched among trans-only predictions (51%; Supplementary Fig. 23b), indicating that distal regulation drives the new findings in our framework. Overlap was observed with SCHEMA61, PGC3-prioritized, Mendelian randomization and fine-mapped gene sets5 (Supplementary Figs. 24–27), and 21 genes were consistently detected in at least four independent studies (Fig. 6c). Although this overlapping set was not enriched for specific gene ontology, key risk loci emerged, including MAPK3 and INO80E, both residing within the 16p11.2 risk locus63, as well as GATAD2A, a component of the NuRD chromatin-remodeling complex, and GNL3, a regulator of neural progenitor maintenance64. Enrichment was strongest for the Gandal gene set58 (P = 2.98 × 10−21) and, to a lesser extent, with the Huckins13 (P = 0.002) and Hall59 (P = 0.023) (Fig. 6d). Mapping coTWAS-significant genes to PGC3 loci5 showed that 158 genes (27.4% of all tissue-specific associations) fell within significant genome-wide regions, while 608 (72.6%) lay outside (Supplementary Fig. 28a). Cis and cis–trans signals predominated within loci, while trans-only associations were largely distal (Supplementary Fig. 28a; average of 2.0% within loci vs 28.2% outside). Several of the strongest associations (−log10(FDR) > 20) clustered outside established GWAS peaks (Supplementary Fig. 28c), highlighting regulatory effects in addition to known SCZ loci. Together, our coTWAS identified 766 SCZ-associated genes, 641 of which represent novel transcriptome-wide associations, demonstrating that incorporating coexpression-informed trans regulation expands gene discovery beyond cis-only TWAS frameworks. Discussion In this study, we developed coexpression-informed trans-eQTL prediction frameworks across six brain regions and integrated them with established cis-based models to improve genetically imputed gene expression. By explicitly modeling distal regulatory effects, we expanded the pool of predictable genes and demonstrated that combining cis and trans components enhances gene-level prediction accuracy. Applying the framework to 102,613 individuals from PGC3, we found 766 significant SCZ-associated genes, 641 of which represent novel transcriptome-wide associations. Our cross-dataset training and validation strategy prioritized reproducible trans signals while preserving the strengths of established cis-based predictors. As expected, cis models explained larger per-gene variance than trans models within the testing dataset (Fig. 3b and Supplementary Fig. 10), reflecting the larger effect size of local eQTLs. However, trans-informed models substantially increased the number of genes for which genetically regulated expression could be robustly imputed (Fig. 3c) across independent datasets (GTEx and CMC; Supplementary Fig. 11). Integration of cis and trans scores significantly enhanced gene-level accuracy relative to single-modality models (Fig. 4a,b), indicating that distal regulatory effects provide complementary information in addition to genetic control. This complementarity aligns with the polygenic and network-based architecture of SCZ, in which multiple interacting genes contribute to multiple biological pathways to influence disease risk28,29,39,65. INGENE and MODULE highlight distinct but complementary features of distal regulation. INGENE models cis-mediated trans effects66 through coexpression partners, whereas MODULE captures shared variance across coexpressed genes28,38. Notably, the two approaches converge on a substantial set of predicted genes, suggesting partially shared regulatory mechanisms (Supplementary Table 1). Consistent with this overlap, up to 29% of MODULE-predictable genes have at least one cis-SNP regulating a coexpression partner used by INGENE, indicating that part of the MODULE signal may reflect cis-mediated effects propagated through coexpression networks. We found that approximately 40% of MODULE trans-eQTLs also function as cis-eQTLs in independent tissues, based on data from the GTEx QTL resource (Supplementary Table 3), consistent with the idea that regulatory variants can exert both local and distal effects. Enrichment of transcriptional regulators among genes linked to trans signals (Supplementary Fig. 13a) further supports a potential regulatory role for transcription factors in mediating the effects of trans-eQTLs on gene expression67,68. Several genes have established links with SCZ-relevant pathways5, including GATAD2A4 (gene silencing), RERE61 (developmental transcriptional repressor), SP4 (associated with SCZ69 and bipolar disorder70) and IRF3 (immune and inflammatory signaling71). The gene ESRRA regulates a DRD2 gene coexpression network associated with SCZ risk40, while the dysregulation of the HDAC1 protein has been widely reported in the disorder72,73. Notably, most MODULE-predicted gene regulatory mechanisms were not mediated by cis effects. The remaining 70% of signals may reflect SNPs acting in cis on upstream regulators without direct transcriptional mediation, context-specific cis-eQTLs not captured in bulk tissue (for example, brain region, cell type or developmental stage specificity) or feedback regulation from expressed genes17,74–76. One explanation for MODULE predictions is that RNA-seq quantification of individual genes is inherently noisy, and incorporating coexpression with related genes reduces such uncertainty. Therefore, coexpression data enhances the detection of functional eQTLs and thus contributes to reducing false negatives. Gene-level association methods11,12 or isoform-level TWAS77 represent a notable advantage by pinpointing genes within loci that fall short of genome-wide significance in GWAS. These studies identified SNPs often exhibiting GWAS P values in a range between 5 × 10−8 and 10−3, suggesting borderline associations that could attain significance in larger studies. Our approach differs from traditional methods by using coexpression-based models, which re-annotate genetic variants to local (cis) and distal (trans) genes based on statistical association and not gene proximity. Notably, SNP weights derived from our trans model MODULE showed stronger positive correlation with PGC3 SCZ GWAS effect sizes than cis-derived weights (Supplementary Figs. 14 and 15), indicating that variants contributing to network-level regulation may be preferentially enriched for disease liability. Gene Ontology enrichment (Fig. 6a) and cell-type specificity analyses (Supplementary Fig. 22) of coTWAS-significant genes highlight two convergent axes of SCZ pathophysiology: dysregulation of AMPA receptor trafficking within excitatory neurons and immune-endosomal activation processes. Consistent with recent large-scale transcriptomic studies78, excitatory neurons exhibit pronounced disease-associated expression changes58. In the DLPFC, vesicular and endocytic genes enriched in excitatory neurons were linked with AMPAR trafficking and receptor internalization dynamics, in line with prior reports of altered receptor trafficking in SCZ79,80. Genes enriched in GABAergic interneurons showed relative downregulation in the same region, supporting models of excitation–inhibition imbalance81,82. Across prefrontal and cingulate regions, enrichment of antigen-processing and MHC pathways aligns with evidence that neuroinflammatory activity exacerbates inhibitory neuron downregulation83. Although classically immune-related57, MHC class I molecules also regulate NMDA and AMPA receptor trafficking and synapse density during development84, suggesting that the glia component may not solely reflect neuroimmune processes. Moreover, given that most associations with expression of immune genes, such as with C4A, were negative, this suggests alternative mechanisms other than increased immune response. Although certain genes are preferentially expressed in specific cell types, the eQTL effects captured here reflect bulk tissue regulation and may integrate contributions from multiple cellular compartments. Comparison of coTWAS hits with prior SCZ studies5,9,13,27,58–61,85 (Fig. 6b–d and Supplementary Fig. 26) demonstrated both replication and novel findings. Although 125 genes were previously reported, over 80% were newly identified (Fig. 6b and Supplementary Fig. 27a), with trans-only models accounting for most novel discoveries (Supplementary Fig. 27b). Significant overlap with Mendelian randomization results5 (Fig. 6b bottom) indicates that a subset of associations is supported by large-scale causal inference analyses, although Mendelian randomization relies exclusively on genome-wide significant cis-eQTL instruments and therefore does not directly validate trans-regulated signals. Several genes replicated across at least four independent SCZ studies (Fig. 6c), converging on pathways involving synaptic signaling, chromatin remodeling and neurodevelopment. Notable examples include MAPK3 (ref. 86), which links glutamatergic transmission to synaptic plasticity programs, and INO80E, GATAD2A and GNL3, previously reported as dysregulated in post-mortem SCZ cortex63,64. Several limitations should be considered. First, predictive model training relied on post-mortem bulk tissue datasets of moderate size, constraining power for trans-eQTL detection and training of prediction models11,87. Although cross-dataset validation mitigates overfitting, larger training cohorts will probably refine effect estimates. Second, bulk tissue data obscure cell-type-specific effects, and integration with large-scale single-cell eQTL resources may improve resolution32,88,89. Third, our study does not account for the influence of sex on gene expression. Sex is known to contribute to the heterogeneity observed in complex diseases like SCZ90,91. Fourth, analyses were restricted to European ancestry, as most individuals in PGC3 have this ancestry. Lastly, as with all imputation-based approaches, associations reflect genetically predicted expression and do not establish causality. Collectively, our findings demonstrate that integrating coexpression as prior information in predictive models and combining cis and trans components improves gene-level modeling in disorders like SCZ, in which trans heritability may contribute substantially to risk. Given that predicted expression is derived independently of GWAS statistics, this framework complements traditional polygenic risk scores while providing gene-level interpretation of genetic liability. The resulting associations provide region-specific directionality of genetic effects, informing potential physiological mechanisms. Overall, our results support a model in which gene–gene interactions within regulatory networks contribute substantially to SCZ risk. Methods Ethics statement The research described herein complies with all relevant ethical regulations. This study analyzed previously generated human genomic and transcriptomic datasets from the LIBD, GTEx, CMC and PGC3. All primary data collection and tissue acquisition were conducted with written informed consent and approved by the relevant Institutional Review Boards as described in the original publications and in the Supplementary Notes. Subjects The training dataset comprised genotype and brain tissues from individuals of European ancestry in the LIBD collection, including neurotypical controls, individuals with SCZ, bipolar disorder and major depressive disorder (Table 1). Replication analyses used GTEx post-mortem data from European ancestry controls aged 20–79 years; CMC post-mortem samples from SCZ, bipolar disorder and neurotypical control European ancestry donors aged 17–90 years (Table 1); and PGC3 genotype data from 102,613 individuals of European ancestry (SCZ = 42%) (Supplementary Table 6). Data sources Selection criteria The discovery dataset included post-mortem brain samples from healthy individuals and individuals diagnosed with SCZ, bipolar disorder or major depressive disorder. To ensure consistency with our primary post-mortem replication dataset (GTEx) and maximize predictive stability, we included only individuals of European ancestry aged ≥17 years. LIBD RNA-seq data Post-mortem tissue homogenate RNA-seq data were obtained from the LIBD collection across DLPFC60,92,93, hippocampus60,94, caudate nucleus95, amygdala93,96, dACC93 and sACC96 (Supplementary Notes). Genotype–RNA concordance was verified following the SPEAQeasy pipeline (https://research.libd.org/SPEAQeasy-example/swap_speaqeasy.htm). Replication RNA-seq data GTEx v8 expression data (dbGaP accession no. phs000424.v8.p2) were used for ACC, amygdala, caudate nucleus, DLPFC (Frontal Cortex BA9) and hippocampus. CMC RNA-seq data were obtained from the CMC (DLPFC, syn18097439 v3.0; ACC, syn29442240 v6.0). Genotype data LIBD genotypes were generated using Illumina BeadChips and imputed through the TOPMed imputation server97 using HRC reference panels98. GTEx genotypes followed the GTEx pilot analysis99 and were imputed with IMPUTE2(ref. 100) using the 1000 Genomes Phase 1 freeze as the reference panel101. CMC samples were genotyped using the Illumina Infinium HumanOmniExpressExome platform and imputed using the HRC reference panel. Data preprocessing All analyses were performed primarily using R statistical software (v.4.5.0). Additional preprocessing and model preparation steps were performed using Python (v.3.10.19). Genotype post-imputation processing LIBD, GTEx and CMC genotype datasets were subjected to uniform post-imputation quality control using PLINK (v.1.09). Variants were removed if they were duplicated, deviated from Hardy–Weinberg equilibrium (P < 10−6) and had a minor allele frequency of <0.01 or missingness of >0.05. Individuals with >2% missing genotypes were excluded. Relatedness was estimated using PLINK --genome, and pairs exceeding third-degree relatedness (identity by descent of >0.125) were removed. Population stratification was evaluated relative to HapMap3 (ref. 102) using principal components; individuals with more than 90% European ancestry were excluded. After quality control, the final SNP counts were 7,521,829 for LIBD, 8,623,182 for GTEx and 5,859,752 for CMC. For downstream analyses, we subset LIBD genotypes to GTEx overlapping SNPs, yielding 6,819,569 SNPs. PGC3 genotype preprocessing We analyzed 62 European ancestry cohorts from PGC3 (Supplementary Table 6). Genotype acquisition, imputation and calculation of genomic eigenvariates for population stratification were performed by the PGC5. Within each cohort, we retained SNPs with INFO > 0.9, missingness of <1%, minor allele frequency of >0.01 and Hardy–Weinberg equilibrium P > 10−6 (PLINK v.1.9). Analysis pipeline Expression quantification and preprocessing Gene-level mRNA expression was quantified using recount (v.3.22)103, with processed counts converted to reads per kilobase per million mapped reads (RPKMs). We selected genes with median RPKM ≥ 0.1 and ≤20% zero values. We log-transformed RPKM in log2(RPKM + 1), and we removed outlier subjects (>3 s.d. by inter-array distance38). After removing mitochondrial genes, the number of retained genes varied by region (Supplementary Table 8), and we converted LIBD and GTEx expression values to transcripts per million. Expression values were residualized separately within each dataset and brain region using linear models adjusting for relevant technical and biological covariates, including diagnosis (where applicable), sex, age, RNA integrity metrics, mapping rates, post-mortem interval (GTEx/CMC), the first five ancestry components and the first three expression principal components. Neuronal proportion, estimated using BRETIGEA104, was strongly correlated with principal components across regions and therefore not included to avoid collinearity. We rank-normalized residuals using RNOmni (v.1.0.1.2)105 before downstream analyses. The number of genes retained per region is shown in Supplementary Table 8. cis-eQTL discovery We defined cis-eQTLs as SNPs located within ±1 Mb of each gene’s transcription start site, defined as the 5′ gene coordinate (start for genes on the positive strand; end for genes on the negative strand). SNP–gene associations were tested using robust linear models (robustbase v.0.99.4.1 (ref. 106)) within a fourfold cross-validated framework. Statistical significance was assessed using F-tests (sfsmisc v.1.1.20 (ref. 107)), and fold-specific estimates were combined by fixed-effect inverse variance meta-analysis (stats v.3.6.2), yielding a single effect size and P value per SNP–gene pair. SNPs were ranked by P value, and linkage disequilibrium pruning was performed within ±250 kb of the lead SNP. Independent signals were defined as R2 < 0.1, with iterative removal of correlated variants38. cis-eQTL-based models training and application We trained our CIS model following the PrediXcan11 method, using SNPs within ±1 Mb of each gene. Elastic-net regression was implemented with the glmnet108 R package (v.4.1.8) and performed in a fourfold cross-validated framework to estimate SNP weights. Models with adjusted R2 ≥ 0.01 (P < 0.05) were retained. Tissue-specific CIS models were applied to test genotypes using MetaXcan109 (Predict.py). To incorporate functional priors, we trained the EpiXcan12 model, which integrates epigenomic annotations into an elastic-net penalty. We retrieved annotations from the Roadmap Epigenomics50 portal and trained tissue-specific models following the published EpiXcan implementation. When specified, we followed the original parameter settings; otherwise, we aligned procedures with the CIS pipeline for consistency. Training and application of coexpression-based predictive models: INGENE and MODULE Coexpression networks We leveraged 48 previously published WGCNA110-derived coexpression networks29,33,34,36–38,58,76,111,112. The gray (unassigned) module was excluded from analysis. The genes and module assignments for all networks used in this study are provided in Supplementary Data 1. INGENE INGENE predicts gene-level expression from genetically inferred expression of cis-regulated coexpression partners. For a given gene g with P partners, the predicted expression Yg is modeled as a weighted linear combination of partner-level predictions, assuming additive genetic effects: where Yg is the predicted expression of gene g, and wp′,g is the estimated effect size of partner p′ on gene g. We first generated cis-based predictions (CIS and EpiXcan) in the LIBD training set. These gene-level predictions served as inputs for coexpression partners within the INGENE framework. When both CIS and EpiXcan predicted the same partner, we retained the predictor with the higher adjusted R2 to maximize partner-level accuracy. For each brain region and each of the 48 coexpression networks, we trained elastic-net regression models (cv.glmnet112 R package) using fourfold cross-validation to estimate partner weights and tune the penalty parameter λ. Models with adjusted R2 ≥ 0.01 (P < 0.05) were retained. Non-zero coefficients were stored to construct the INGENE prediction database. To apply INGENE to the testing data, we first imputed cis-mediated partner expression using CIS and EpiXcan weights, then applied INGENE-derived coefficients to generate gene-level predictions. For genes predicted by multiple networks, we averaged predictions and selected the cis-derived estimate with the highest adjusted R2 between predicted and observed expression for inclusion in the final INGENE output. MODULE The MODULE framework predicts gene expression using candidate trans-eQTLs associated with the coexpression module eigengene. For each network and brain region, we computed the module eigengene as the first principal component (PC1) of the region-specific coexpression matrix. For a gene g with a set of k trans markers associated with the module eigengene (excluding its cis-SNPs within ±1 Mb), the trans component of predicted expression is given by: where Yg is the predicted expression of gene g, k−cis is the set of markers removing cis-SNPs, wk−cis,g is the SNP effect size and Xk−cis is the allele dosage. To identify SNPs associated with the module eigengene, we first mapped European 1000 Genomes SNPs113 within ±100 kb of module genes using MAGMA (v.1.09b)113. We then performed fourfold cross-validation (stratified by diagnosis), computing PC1 within each training fold. For each SNP, we tested genotype–PC1 associations using robust linear models (robustbase106 R package) and assessed significance with F-tests (sfsmisc107 R package). SNPs were ranked by P value, and cross-fold consistency was summarized using the rank product. Linkage disequilibrium pruning was performed within ±250 kb of the lead SNP (R2 < 0.1), iteratively removing lower-ranked correlated variants38. SNPs within the top 5% of the rank distribution were retained as co-eQTLs. LIBD genotypes were restricted to selected co-eQTLs, excluding SNPs within ±1 Mb of each target gene to remove residual cis effects. Elastic-net models were trained with nested cross-validation to tune λ (Supplementary Notes). Genes with adjusted R2 ≥ 0.01 (P < 0.05) were retained, and non-zero weights were stored using RSQLite (v.2.4.7)114. We applied models to test genotypes using MetaXcan109. As described in the INGENE section, we used an averaging strategy to consolidate predictions, resulting in a single imputed expression value for each gene. Cross-dataset training of INGENE and MODULE To enhance robustness and minimize dataset-specific effects, trans models were trained independently in LIBD, GTEx and CMC for each available brain region. All trained models were then applied to LIBD genotypes to generate predictions on a common individual-level reference panel. For each gene g, this approach yielded three predicted-expression vectors over the same LIBD individuals: Replicability was assessed by model-to-model agreement rather than prediction-to-observed expression. For each gene, we computed Pearson correlations between LIBD-trained predictions and those derived from GTEx or CMC: Genes were retained if at least one cross-dataset correlation was positive (r > 0). This filtering step was applied to LIBD-trained models and did not incorporate testing datasets in predictive score generation, thereby avoiding information leakage. Functional and regulomic analysis of trans-regulated genes From each tissue-specific MODULE model, we selected trans-SNPs predicting genes validated in GTEx and retained trans-SNPs with absolute predictive weights of ≥0.01. We focused on SNPs that were also significant cis-eQTLs across all 49 GTEx tissues53. SNPs were stratified into quartiles based on absolute MODULE weights and linked to their cis-eGenes (GTEx catalog; Supplementary Table 3). Gene Ontology enrichment was performed separately by brain region and for the highest-weight quartile using the ClusterProfiler package115 (v.3.22), restricting analyses to molecular function. The universe was defined as all unique GTEx genes. For regulomic analysis, we used the gProfileR package116 (v.0.2.4) to identify transcription factors enriched among GTEx cis-eGenes linked to MODULE trans-SNPs. The background was defined as genes regulated in trans within the MODULE framework. Significance was assessed using hypergeometric testing (stats R package v.3.6.2), with FDR and Bonferroni corrections applied for multiple comparisons. Correlation between predictive SNP weights and PGC3 odds ratios We obtained GWAS summary statistics for SCZ, bipolar disorder and major depressive disorder from the PGC European ancestry meta-analysis (links in Supplementary Table 9). For SNPs predicting genes validated in GTEx, we computed mean absolute elastic-net weights from CIS, EpiXcan and MODULE models. GWAS summary statistics were restricted to SNPs with P < 0.05 and overlapping model-specific SNP sets (Supplementary Table 4). We assessed the correlation between SNP weights and log(odds ratio) using Pearson correlation (two-tailed). Correlation coefficients were transformed using Fisher’s z to stabilize variance. Differences between models were evaluated using: where standard errors were derived from the number of SNPs contributing to each model. Two-tailed P values were obtained from the standard normal distribution. Model evaluation ANOVA analysis in GTEx We assessed whether combining cis and trans predictions improved explanatory power using maximum likelihood estimation (car R package117 v.3.1-5). For each brain region, observed GTEx expression (log transcripts per million) was modeled with covariates including sex, mean age, RNA integrity number, RNA rate, overall mapping rate, post-mortem interval, the first three principal components and the first five genomic eigenvariates. For genes predicted by a single model (CIS (C), EpiXcan (E), INGENE (I) or MODULE (M)), we compared: For genes with both cis and trans components, we compared: For trans-only genes: The model (cis-only, cis–trans or trans-only) yielding the most significant improvement in fit was selected for each gene (maximum likelihood estimation, α = 0.05). coTWAS analysis in PGC3 cohorts Gene prediction and selection For each PGC site, we imputed gene expression using tissue-specific CIS, EpiXcan, INGENE and MODULE models. When a gene was predicted by multiple networks, we averaged predictions. We retained only genes validated in GTEx (adjusted R2 and Pearson’s r > 0). Genes present in only one model were classified as unimodal; those included in at least two were classified as multimodal. Association testing Within each PGC cohort, we performed gene-wise logistic regression with diagnosis as the outcome and imputed expression as the predictor, covarying for sex and genomic eigenvariates previously associated with diagnosis5. Unimodal genes were tested using model-specific predictions (equation 1). Multimodal genes were tested using the combined cis–trans predictor trained in GTEx and validated in CMC (equation 2). To ensure cross-cohort consistency, we focused on gene predictors available in ≥32 PGC3 sites. Site-specific β from logistic regression were meta-analyzed using the Stouffer method (metap v.1.12 (ref. 118)), applying weights proportional to the square root of the sample size for each cohort119. To assess heterogeneity, we computed Cochran’s Q-test and retained genes with a P value threshold of >1 × 10−3 across PGC3 sites. In total, 96,535 tissue–gene tests were evaluated. Conditional analysis To identify independent associations and control for correlation among predicted genes arising from linkage disequilibrium between model SNPs, we performed a conditional analysis following GCTA-COJO theory120 using CoCo, an R implementation of GCTA-COJO as previously implemented13. We defined 315 genomic regions (1 Mb window), including the MHC, based on the 96,535 meta-analyzed tissue–gene pairs. Within each region, we generated a predicted expression correlation matrix for all tissue–gene pairs by aggregating cohort-specific correlations using root-effective sample size (neff, equation (3) weighting, restricted to cohorts with available predictions. Forward stepwise joint regression was applied to all tissue–gene pairs passing collinearity filtering (r < 0.9). Genes were iteratively added based on the meta-analytic P value until no additional gene met the significance threshold or no joint statistic reached the significance threshold. We corrected the final output list for genome-wide multiple comparisons using the FDR at α = 0.01. We calculated effect sizes and odds ratios for SCZ-associated genes by adjusting CoCo betas to have unit variance (equation 4): where GVAR is the pooled neff weighted variance of the genetically predicted expression of each gene across all cohorts. coTWAS replicability analysis Given that we did not have access to a PGC-independent dataset, we performed leave-one-site-out validation across the 62 PGC cohorts for gene–tissue pairs passing conditional FDR ≤ 0.01. For each gene, we evaluated directional concordance, defined as the proportion of sites in which the β of the logistic regression fitted on all sites but one has the same direction (same sign) as the β obtained from the model on the held-out site, and the R2 of the correlation between leave-one-out and held-out β estimates across sites. Statistical significance was assessed by permuting held-out β values 10,000 times to generate null distributions for both metrics. Replicable genes were defined as those exceeding the 95th percentile (one-tailed α = 0.05) for both directional concordance and R2. Cell-type specificity analysis We used cell-type specificity indices derived from single-nucleus RNA-seq of human brains54, which distinguish between ten different cell types (neurons and glia). Enrichment was assessed using the mean-rank gene set test implemented in limma (v.3.46)121. Multiple comparisons across genes and cell types were controlled using FDR (α = 0.05). Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Online content Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41588-026-02646-3. Supplementary information Supplementary Notes and Supplementary Figs. 1–28 (Supplementary Tables 1–9 provided separately as an Excel file). Reporting Summary Peer Review File Supplementary Tables 1–9. Excel workbook containing the coexpression network data used in this study. The first worksheet lists the reference and source for each network included in the analyses, and subsequent worksheets provide the genes and module assignments for each network. Acknowledgements F.R. and D.F. are PhD students enrolled in the National PhD in Artificial Intelligence (XXXVII and XXXVIII cycle, respectively) course on Health and Life Sciences, organized by Università Campus Bio-Medico di Roma. We thank L. Fazio and G. Leggio for their scientific contribution to grants that funded this work. We thank S. Han for his helpful suggestions and review of the manuscript. We thank P. Provero for his scientific insight and review of the manuscript. We are grateful for the contributions of the Office of the Chief Medical Examiner of the State of Maryland, Office of the Chief Medical Examiner of Kalamazoo County Michigan, Office of the Chief Medical Examiner University of North Dakota School of Medicine, Gift of Life of Michigan, Office of the Chief Medical Examiner of Santa Clara County, California, and Medical University of Sofia, Bulgaria, in assisting the LIBD in the acquisition and curation of brain tissue donations for this study. All research at the LIBD is made possible by generous gifts from the families of Steve and Connie Lieber and Milton and Tamar Maltz. We would like to thank all the family members of the donors for their exceptional contribution. Special recognition goes to M. Parihar and N. Penzel for providing scientific and technical support. Author contributions G.P. and F.R. conceived and designed the study. D.R.W., T.M.H., J.E.K. and M.C.O. curated the data. F.R., L.S., G.C.K., G.G., F.D.C., D.F., C.J.B. and D.M. performed formal analysis. G.P., D.R.W., S.P. and B.V. acquired funding. F.R. and G.P. conducted the investigation. G.P., F.R., L.S., G.C.K., D.F. and D.M. developed the methodology. G.P. was responsible for project administration. D.R.W., G.B., T.M.H., J.E.K., A.R., S.R., A. Braun, J.K., S.I.B., P.R.M., C.A., J.T.R.W., M.C.O., M.J.O., D.B., A.C., D.W.M., E.D., J.v.O., E.A., M.C.S., M.D.F., B.T.B., C.N.P., A.M., V.G., N.K., V.E.P., A.G., E.K., J.A.C., M. Rivera, D.C., C.L.L., A.S., O.A.A., D.S.C., T.L., A.K.M., N.S.M., B.J.M., D.R., I.G., A.H., B.K., M.M.N., M. Rietschel, G.K., P.F.S., T.L.P., T.W., A.M.M., T.E., E.G.J., H.E., B.P.R., D.F.L., J.D.B., E.B., C.M.H., R.A.O., R.A., E.A.S., B.P.F.R., S.G., F.T. and C.D. acquired resources. F.R. implemented the software. G.P. supervised the project. F.R., L.S., G.C.K. and D.F. prepared visualizations. F.R., G.P. and D.R.W. wrote the original draft of the manuscript; all authors contributed to review and editing. Peer review Peer review information Nature Genetics thanks Chunyu Liu and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available. Funding This work was supported by the University of Pisa through Fondo di Finanziamento Ordinario (FFO), supporting the PhD scholarship of F.R. (“Identifying biological pathways of psychiatric risk from genes to cognition via AI”). B.V., S.P. and G.P. disclose support for the research of this work from the Italian Ministry of University and Research PRIN 2020 project “Hot for genes—the role of brain gene expression in identifying developmental trajectories and malleable risk factors for preventive interventions” (Prot. 2020WSCSLZ). G.P. and S.P. disclose support for the research of this work from the Italian Ministry of University and Research PRIN 2022 project “A prosocial training program to harness genes and environment influencing social behavior” (Prot. 2022KXJYJA). G.P. discloses support for the research of this work from the Italian Ministry of University and Research PRIN PNRR 2022 project “Unraveling new neural network activities for the treatment of negative symptoms and socio-cognitive abilities in schizophrenia” (Prot. P2022HNBJX). The LIBD supported the collection and analysis of post-mortem brain tissue.Open access funding provided by Università degli Studi di Bari Aldo Moro within the CRUI-CARE Agreement. Data availability All data supporting the conclusions of this study are available within the article and its Supplementary Information. LIBD post-mortem RNA-seq data from the BrainSeq Phase I–III studies for caudate nucleus, DLPFC and hippocampus are available through the Database of Genotypes and Phenotypes (dbGaP) and Globus collections (caudate nucleus, phs003495.v1.p1, https://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs003495.v1.p1; DLPFC, jhpce#bsp2-dlpfc, http://research.libd.org/globus/jhpce_bsp2-dlpfc/index.html; hippocampus, jhpce#bsp2-hippo, http://research.libd.org/globus/jhpce_bsp2-hippo/index.html). LIBD genotype data are available through dbGaP under accession phs000979.v3.p2 (https://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs000979.v3.p2). RNA-seq data for amygdala and sACC from the BipSeq study are available through the PsychENCODE Consortium through Synapse (syn5844980) and are managed by the National Institute of Mental Health Repository and Genomics Resource. Raw and processed data for DLPFC, amygdala and dACC samples from the Department of Veterans Affairs Posttraumatic Stress Disorder study are available upon request through the PTSD Brain Bank Resource Request process (https://www.research.va.gov/programs/tissue_banking/ptsd/default.cfm). GTEx post-mortem processed RNA-seq data are publicly available through the GTEx Portal (release v.8). Individual genotype data are available with controlled access through dbGaP under study accession phs000424.v8.p2. CMC MSSM-PENN-Pitt RNA-seq data were obtained upon request through Synapse (DLPFC release 3.0, syn18097439; ACC release 6.0, syn29442240). Corresponding genotype data were accessed at https://www.synapse.org/Synapse:syn18097441. PGC3 individual-level phenotype and genotype data were obtained upon request and are subject to controlled access. Access details can be found at https://pgc.unc.edu/for-researchers/data-access-committee/data-access-information. Researchers must apply for access and comply with the data use agreement. The GWAS summary statistics are publicly available from the PGC for SCZ, bipolar disorder and major depressive disorder through Figshare repositories. GTEx cis-eQTLs data for 59 tissues are publicly available at https://gtexportal.org/home/downloads/adult-gtex/qtl. The study also used previously published WGCNA coexpression networks as detailed in the article and in Supplementary Table 9. Coexpression networks are also provided as Supplementary Data 1 to facilitate reproducibility. Gene expression prediction models and analysis outputs generated in this study are available in the Zenodo repository associated with this work (10.5281/zenodo.14959416)122. These include prediction models trained in the LIBD dataset (MODULE, INGENE, CIS and EpiXcan) across six brain regions, combined cis–trans prediction models validated in GTEx and coTWAS association results, including cross-validation outputs. Code availability Custom code developed for the coTWAS analysis pipeline is publicly available at Zenodo (10.5281/zenodo.14959416)122. The actively maintained version is hosted on GitHub (https://github.com/fabiana1rossi/coTWAS). This study also used publicly available software, including MetaXcan and PrediXcan (https://github.com/hakyimlab/MetaXcan) and EpiXcan (https://bitbucket.org/roussoslab/epixcan/src/master). A complete list of software packages, tools and resources is provided in Supplementary Table 9. Competing interests A. Bertolino received consulting fees from Biogen and lecture fees from Otsuka, Janssen and Lundbeck. D.R.W. serves on the scientific advisory board of Pasithea Therapeutics. G.P., G.B., A.R., F.D.C. and G.C.K. received lecture fees from Lundbeck. A.K.M. is a consultant to Genomind, InformedDNA and Concert Pharmaceuticals. M.C.O., M.J.O. and J.T.R.W. are supported by collaborative research grants from Takeda Pharmaceuticals. O.A.A. is a consultant for HealthLytix and received speaker’s honoraria from Lundbeck. C.A. has been a consultant to or has received honoraria or grants from Acadia, Angelini, Gedeon Richter, Janssen Cilag, Lundbeck, Minerva, Otsuka, Roche, Sage, Servier, Shire, Schering Plough, Sumitomo Dainippon Pharma, Sunovion and Takeda. E.A.S. holds stocks in Regeneron Pharmaceuticals Inc. The other authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary information The online version contains supplementary material available at 10.1038/s41588-026-02646-3. References Associated Data Supplementary Materials Supplementary Notes and Supplementary Figs. 1–28 (Supplementary Tables 1–9 provided separately as an Excel file). Reporting Summary Peer Review File Supplementary Tables 1–9. Excel workbook containing the coexpression network data used in this study. The first worksheet lists the reference and source for each network included in the analyses, and subsequent worksheets provide the genes and module assignments for each network. Data Availability Statement All data supporting the conclusions of this study are available within the article and its Supplementary Information. LIBD post-mortem RNA-seq data from the BrainSeq Phase I–III studies for caudate nucleus, DLPFC and hippocampus are available through the Database of Genotypes and Phenotypes (dbGaP) and Globus collections (caudate nucleus, phs003495.v1.p1, https://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs003495.v1.p1; DLPFC, jhpce#bsp2-dlpfc, http://research.libd.org/globus/jhpce_bsp2-dlpfc/index.html; hippocampus, jhpce#bsp2-hippo, http://research.libd.org/globus/jhpce_bsp2-hippo/index.html). LIBD genotype data are available through dbGaP under accession phs000979.v3.p2 (https://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs000979.v3.p2). RNA-seq data for amygdala and sACC from the BipSeq study are available through the PsychENCODE Consortium through Synapse (syn5844980) and are managed by the National Institute of Mental Health Repository and Genomics Resource. Raw and processed data for DLPFC, amygdala and dACC samples from the Department of Veterans Affairs Posttraumatic Stress Disorder study are available upon request through the PTSD Brain Bank Resource Request process (https://www.research.va.gov/programs/tissue_banking/ptsd/default.cfm). GTEx post-mortem processed RNA-seq data are publicly available through the GTEx Portal (release v.8). Individual genotype data are available with controlled access through dbGaP under study accession phs000424.v8.p2. CMC MSSM-PENN-Pitt RNA-seq data were obtained upon request through Synapse (DLPFC release 3.0, syn18097439; ACC release 6.0, syn29442240). Corresponding genotype data were accessed at https://www.synapse.org/Synapse:syn18097441. PGC3 individual-level phenotype and genotype data were obtained upon request and are subject to controlled access. Access details can be found at https://pgc.unc.edu/for-researchers/data-access-committee/data-access-information. Researchers must apply for access and comply with the data use agreement. The GWAS summary statistics are publicly available from the PGC for SCZ, bipolar disorder and major depressive disorder through Figshare repositories. GTEx cis-eQTLs data for 59 tissues are publicly available at https://gtexportal.org/home/downloads/adult-gtex/qtl. The study also used previously published WGCNA coexpression networks as detailed in the article and in Supplementary Table 9. Coexpression networks are also provided as Supplementary Data 1 to facilitate reproducibility. Gene expression prediction models and analysis outputs generated in this study are available in the Zenodo repository associated with this work (10.5281/zenodo.14959416)122. These include prediction models trained in the LIBD dataset (MODULE, INGENE, CIS and EpiXcan) across six brain regions, combined cis–trans prediction models validated in GTEx and coTWAS association results, including cross-validation outputs. Custom code developed for the coTWAS analysis pipeline is publicly available at Zenodo (10.5281/zenodo.14959416)122. The actively maintained version is hosted on GitHub (https://github.com/fabiana1rossi/coTWAS). This study also used publicly available software, including MetaXcan and PrediXcan (https://github.com/hakyimlab/MetaXcan) and EpiXcan (https://bitbucket.org/roussoslab/epixcan/src/master). A complete list of software packages, tools and resources is provided in Supplementary Table 9.