TrASPr

TrASPr (Transformer for Alternative Splicing Prediction) is a deep learning model for predicting tissue-specific alternative splicing of pre-mRNA. Alternative splicing lets a single gene produce multiple transcript isoforms in a context-dependent way, and defects in this process underlie many diseases. Accurately predicting how much a given exon is included in a given tissue — and which sequence elements drive that decision — is therefore central to both basic regulatory biology and RNA therapeutics.

Developed by Wu, Barash, and colleagues at the University of Pennsylvania and released as a preprint in January 2025, TrASPr is the predictive core of the TrASPr+BOS framework. It is a multi-transformer model that ingests genomic sequence around splice sites together with tissue identity and other contextual features to predict percent-spliced-in (PSI) values and their differences across conditions. A distinguishing property is its ability to generalize zero-shot to cellular conditions not seen during training, rather than being locked to the tissues in its training set.

In the full framework, TrASPr also acts as an oracle to label candidate sequences for a companion Bayesian Optimization for Splicing (BOS) routine that designs RNA for a chosen splicing outcome; because that optimization loop is run per design target, this entry focuses on TrASPr as a general-purpose splicing predictor.

#Key Features

• Tissue-specific PSI prediction: Predicts percent-spliced-in values and differential splicing (dPSI) across tissues, capturing how the same exon is regulated differently by cellular context.
• Multi-transformer architecture: Uses several transformer modules operating on the sequence around the relevant splice sites, allowing it to handle different classes of alternative splicing events within one model.
• Zero-shot generalization: Extends to cellular conditions not present in training, improving tissue-specific AUPRC by up to 1.8-fold over prior methods.
• Interpretable regulatory signal: Recovers tissue-specific regulatory elements, and predicted splicing changes were experimentally probed, including validation of regulatory elements using dCas13.

#Technical Details

TrASPr processes 6-mer-tokenized genomic sequence windows centered on splice sites through a multi-transformer backbone, combining these representations with tissue tokens, exon and intron length features, and evolutionary conservation values before predicting PSI and differential-PSI targets. The released implementation trains across six tissue types — lung, heart, brain, liver, spleen, and lymphocytes — using pre-training on splice-site-centered sequences followed by fine-tuning on labeled PSI data. The model reports up to a 1.8-fold improvement in tissue-specific AUPRC relative to existing splicing predictors, and the authors validated hundreds of predicted novel tissue-specific splicing variations experimentally. Code and example configurations are available on GitHub, and the preprint is released under a CC BY license.

#Applications

TrASPr is useful to RNA biologists and therapeutic developers who need to anticipate how sequence context or genetic variants will shift splicing in a particular tissue. Typical uses include screening variants for splicing effects, mapping the cis-regulatory elements that control exon inclusion, and prioritizing candidate targets for splice-modulating interventions. As the predictive oracle inside TrASPr+BOS, it also underpins model-guided design of RNA sequences intended to achieve a specified condition-specific splicing outcome.

#Impact

TrASPr advances splicing prediction by combining a multi-transformer design with demonstrated zero-shot transfer to unseen conditions and experimental validation of its predictions, addressing a long-standing limitation of tissue-specific generalization. By coupling accurate prediction with an optimization layer for sequence design, the broader framework points toward programmable control of splicing for research and therapeutic applications. As a preprint, its influence will grow as the community benchmarks and applies it to new tissues and variants.