How AI is Organising Alternative Splicing Prediction to Target Disease
Source PublicationeLife
Primary AuthorsWu, Maus, Jha et al.
Inside every human cell, a silent, microscopic editing process determines our survival. A single genetic sequence can be sliced and pasted in thousands of different ways to build entirely different proteins. When this editing system fails, the resulting cellular errors can trigger devastating neurological diseases and cancers.
These results were observed under controlled laboratory conditions, so real-world performance may differ.
To address this, researchers have developed a generative artificial intelligence system called TrASPr+BOS. The system uses a multi-transformer model to forecast how genes edit themselves across different human organs, outperforming previous computational methods by up to 1.8-fold in accuracy. This AI acts as an artificial oracle, generating synthetic data to train an optimisation algorithm that can design custom RNA sequences for specific tissues.
The Next Era of Alternative Splicing Prediction
The system achieves this through three distinct capabilities:
- Generalising to unseen cellular conditions across different human organs.
- Generating high-fidelity synthetic data to train optimisation algorithms.
- Identifying novel regulatory elements that can be validated in the laboratory.
The team validated hundreds of novel, tissue-specific variations and confirmed these new regulatory elements using molecular tools in the laboratory. The model measured a significant increase in prediction accuracy across unseen cellular conditions, demonstrating its ability to generalise beyond its training data.
These findings suggest that researchers may soon design bespoke RNA therapies to correct splicing errors in specific organs, such as the brain, without affecting healthy tissue. This approach could lead to highly targeted treatments for genetic disorders that were once deemed untreatable.