AI model designs RNA splicing from scratch
A new AI model can predict how RNA is spliced in different tissues, and can also design new RNA sequences that produce a desired splicing outcome.
Genes contain more information than cells actually use. After DNA is copied into RNA, sections are cut out and the remaining pieces are joined together. This process is called alternative splicing. Which sections are kept differs by tissue and cell type. In some diseases, this process goes wrong, producing faulty proteins.
The researchers built a model called TrASPr+BOS that combines a language model for RNA sequences with an optimisation algorithm. It does not just predict what happens to a given sequence; it can also design new sequences that produce a specific splicing pattern. That generative step is new.
From prediction to design
Earlier models could predict how a given RNA sequence would be spliced with reasonable accuracy. But moving from prediction to design is a significant leap. It requires the model to understand which parts of a sequence determine the outcome, and to generate new sequences meeting specified criteria.
The model learned these patterns from large datasets of splicing events in human tissue. It can now produce sequences that splice in a desired way in one tissue while behaving differently in others. That tissue specificity is exactly what targeted therapies require.
Relevance to disease and aging
Disrupted splicing contributes to cancer, muscle diseases, and neurodegenerative conditions. In aging, splicing patterns shift over time as tissues increasingly produce abnormal proteins through altered splicing. A model that can predict and correct this is directly applicable.
The method has not yet been tested clinically. But the combination of prediction and generative design in a single model is a step previous methods lacked. It provides a concrete basis for therapeutic RNA design.