Motivation for isotwas

isotwas provides functions to run transcriptome-wide association studies on the isoform-level.

Isoform-level analyses can provide further granularity to a gene-trait association by pinpointing the isoform of a given gene that drives the association. Furthermore, if two isoforms of the same genes have associations with divergent effect sizes, then gene-level trait mapping will likely miss this association but isoform-level trait mapping will not.

isotwas contains functions for:

1. Training multivariate predictive models for isoform-level expression from genetic variants, with 9 different methods including:

   • MRCE (Multivariate Regression with Covariance Estimation)
   • Multivariate elastic net
   • Joinet stacked elastic net
   • Sparse PLS
   • Sparse group lasso
   • Multi-task lasso (L21 regularization)
   • Super learner stacking
   • Graph-regularized regression (using isoform similarity)
   • Univariate methods (elastic net, BLUP, SuSiE)

2. Conducting trait mapping on the isoform-level using stage-wise hypothesis testing

3. Probabilistic fine-mapping to identify causal isoforms

Key Features

• Multiple prediction methods: The package automatically evaluates multiple multivariate regression methods and selects the best model based on cross-validated R-squared.

• Graph-regularized learning: Incorporate prior knowledge about isoform similarity (e.g., shared exon structure) to improve predictions.

• User-friendly interface: The main compute_isotwas() function provides detailed progress output, data summaries, and results comparison tables.

• Robust model selection: Uses cross-validation within each method and compares across methods to find the optimal model for each gene.

See the Train an isoTWAS model vignette for detailed usage instructions.