Genome-wide association studies (GWAS) have identified numerous genetic associations with complex traits. However, translating these associations into molecular mechanisms remains challenging, especially as GWAS hits often localize to non-coding regions within large linkage disequilibrium (LD) blocks. Prioritizing causal variants and identifying their functional impact on target genes requires innovative methods. Transcriptome-wide association studies (TWAS) integrate population-level transcriptomic reference panels with GWAS summary statistics to prioritize genes, but they largely focus on total gene expression, ignoring the contribution of distinct transcript isoforms generated through alternative splicing. Alternative splicing is a tissue-specific gene regulatory mechanism affecting ~90% of human genes, significantly expanding the genome's coding and regulatory potential. The brain, in particular, exhibits exceptionally complex splicing patterns, contributing to its evolutionary and phenotypic complexity. Brain-expressed genes are longer, contain more exons, and show more complex splicing than genes in other tissues. Independent of gene expression, splicing dysregulation has been implicated in disease, especially neuropsychiatric disorders. While isoform-level expression changes show greater enrichment for schizophrenia heritability than gene or local splicing changes, integrating these quantifications with GWAS requires novel computational methods capable of handling the high correlation between isoforms of the same gene. This study addresses this gap by introducing a novel approach.

Numerous methods have been developed to integrate population-level transcriptomic reference panels with GWAS summary statistics to prioritize genes at trait-associated loci. These methods, including transcriptome-wide association studies (TWAS), impute the *cis*-component of gene expression predicted by common variants into an association cohort, reducing multiple comparisons and increasing interpretability. However, previous integrative analyses have largely focused on total gene expression and not on the distinct transcript isoforms of a gene. Alternative splicing, a tissue-specific mechanism, vastly expands the genome's coding and regulatory potential and is particularly relevant in the brain. Studies have shown that isoform-level expression changes have greater enrichment for schizophrenia heritability than gene or local splicing changes. However, computational methods capable of jointly modeling highly correlated isoforms were lacking, highlighting the need for methodological innovation in this area.

The authors introduce isoTWAS, a three-step isoform-level TWAS framework. First, multivariate predictive models of isoform-level expression are built from SNPs within 1 Mb in functional genomics training datasets (GTEx and PsychENCODE). Four multivariate penalized predictive frameworks (multivariate elastic net, multivariate LASSO regression with covariance estimation, multivariate elastic net with stacked generalization, and sparse partial least squares) and a univariate approach (univariate elastic net, BLUP, SuSiE) are compared, selecting the best-performing model based on 5-fold cross-validation (CV) R². Second, these models are used to impute isoform expression into an external GWAS cohort and quantify the association with the target GWAS phenotype. A weighted burden test is used if only GWAS summary statistics are available. Third, isoTWAS performs a stepwise hypothesis-testing procedure to account for multiple comparisons and local LD structure. Isoform-level P-values are aggregated to the gene-level using the aggregated Cauchy association test (ACAT) with false discovery rate control. Isoforms of prioritized genes undergo post-hoc family-wise error controls. A permutation test assesses the signal added by isoform expression, controlling for large LD blocks. Optional isoform-level Bayesian fine mapping identifies the minimal credible set of isoforms. The isoTWAS framework is available as an R package. The authors also describe their simulation framework to evaluate isoTWAS performance, generating isoform expression with specified isoQTL architecture and controlled heritability. They simulated scenarios with only gene-level effects, a single effect isoform, and two effect isoforms with varying magnitudes of association. Prediction performance is assessed using multiple criteria, including the number of imputed isoforms and genes, and the accuracy of total gene expression prediction, which improved substantially with isoTWAS across simulations and real data from GTEx in 48 tissues (13 brain tissues). The accuracy of prediction is related to detection power in the trait mapping step. The performance is further evaluated in the context of 15 neuropsychiatric traits.

isoTWAS significantly outperformed gene-level TWAS in both simulated and real data. In simulations, isoTWAS showed improved isoform and gene expression prediction accuracy across various genetic architectures, particularly when isoQTLs were sparse or effects diverged across isoforms. In GTEx data, isoTWAS trained 2.3-2.5 times more models than univariate approaches and improved predictions for 79-82% of isoforms, with a median increase of 1.8-2.4 fold in adjusted R². The improvements were particularly pronounced in brain tissues, suggesting more shared isoQTL architecture in the brain. isoTWAS also increased the number of testable genes and improved total gene expression prediction by a median of 25-70% in CV, exceeding TWAS's performance in out-of-sample predictions using PsychENCODE/AMP-AD data. Applying isoTWAS to 15 neuropsychiatric GWAS datasets, the study found that isoTWAS identified more gene-trait associations than TWAS across both adult and developmental prefrontal cortex reference panels (adult: 2595 vs 1589; developmental: 4062 vs 890). isoTWAS detected 60% more associations at GWAS loci compared to TWAS, prioritizing genes within more loci in schizophrenia analysis. Moreover, isoTWAS findings were well-calibrated to the null and did not show significant inflation. isoTWAS-specific associations were enriched for constrained genes, indicating increased identification of true disease signals. Fine-mapping analysis using isoTWAS and TWAS showed comparable results, highlighting the potential for improved resolution in the future. The superior performance of isoTWAS over a splicing-event-based approach emphasizes the unique advantages of its isoform-centric focus. The study also highlighted specific cases where isoTWAS detected trait associations undetectable by TWAS, implicating isoforms of AKT3, CUL3, HSPD1, and PCLO in genetic associations with psychiatric traits. These isoTWAS-prioritized genes were enriched for pathways consistent with the biology of the underlying traits. The findings demonstrate that isoTWAS effectively prioritizes novel candidate risk mechanisms by incorporating isoform-level regulation.

The study demonstrates that incorporating isoform-level resolution within integrative genomic approaches significantly increases the discovery of trait associations, especially for brain-related traits. isoTWAS's superiority over gene-level methods is attributed to its ability to capture isoQTL architectures that vary across isoforms, improved prediction accuracy due to multivariate modeling, and increased power in detecting trait associations by leveraging the increased number of testable genes and improved gene-level predictions. The isoform-centric approach offers a more granular view of disease mechanisms, addressing limitations of traditional gene-level analyses. While splicing-event-based methods offer insights, isoTWAS shows superior performance, as its focus on isoforms captures the combined effects of multiple splicing events. The study acknowledges limitations such as dependence on transcriptome annotations and short-read RNA-seq limitations; the advent of long-read sequencing will enhance annotation and improve isoform quantifications. Other areas for future development include incorporating technical variation measures and extending the framework to analyze genetically regulated transcript usage and accounting for horizontal pleiotropy.

isoTWAS provides a powerful framework for identifying isoform-level mechanisms underlying genetic associations with complex traits. The superior performance of isoTWAS compared to gene-level and splicing-event based TWAS highlights the need to shift focus to isoform-level quantifications to maximize discovery. Future work should focus on improving isoform quantifications, addressing methodological challenges, and incorporating features such as horizontal pleiotropy to further enhance the power and accuracy of isoTWAS.

The study acknowledges several limitations. First, isoform-level expression quantifications rely on maximum likelihood estimates from short-read RNA-seq and are dependent on the accuracy and completeness of transcriptome annotations. Second, technical variation from RNA-seq quantification is not incorporated into the predictive models. Third, while isoTWAS can be extended to analyze genetically regulated transcript usage, this requires further methodological development. Lastly, horizontal pleiotropy can reduce isoTWAS power and fine-mapping sensitivity, necessitating further methodological improvements.