Generative modeling for RNA splicing prediction and design

Area of Science:

• Molecular Biology
• Computational Biology
• Genomics

Background:

• Alternative splicing (AS) is vital for tissue-specific gene regulation.
• Splicing defects are linked to various diseases.
• Predicting and manipulating AS offers therapeutic potential.

Purpose of the Study:

• To introduce TrASPr+BOS, a novel AI model for predicting and designing RNA with specific tissue-based splicing outcomes.
• To leverage generative AI and Bayesian Optimization for enhanced AS control.

Main Methods:

• Developed TrASPr (Transformer for Alternative Splicing Prediction), a multi-transformer model adaptable to diverse AS events and cellular conditions.
• Integrated TrASPr as an oracle to generate data for training a Bayesian Optimization for Splicing (BOS) algorithm.
• Utilized dCas13 for experimental validation of predicted splicing variations and regulatory elements.

Main Results:

• TrASPr+BOS demonstrated superior performance compared to existing methods.
• Achieved up to a 1.8-fold enhancement in tissue-specific AUPRC (Area Under the Precision-Recall Curve).
• Successfully identified and validated novel tissue-specific splicing variations and regulatory elements.

Conclusions:

• TrASPr+BOS provides an accurate and efficient approach for AS prediction and RNA design.
• The model advances the understanding of tissue-specific gene regulation.
• Offers a valuable tool for researchers in molecular biology and therapeutic development.