Advances and challenges of splicing prediction with AI

Area of Science:

• Genomics and Molecular Biology
• Computational Biology and Bioinformatics
• Genetics and Disease

Background:

• RNA alternative splicing is a critical post-transcriptional process.
• Dysregulation of splicing is implicated in numerous human diseases.
• Accurate splicing outcome prediction is vital for precision medicine.

Purpose of the Study:

• To review the evolution of computational approaches for predicting RNA alternative splicing.
• To analyze factors influencing predictive performance, including data scale, resolution, quantification, and model complexity.
• To discuss the translational applications and persistent challenges in splicing prediction.

Main Methods:

• Review of computational methodologies, from statistical heuristics to artificial intelligence frameworks.
• Dissection of key factors affecting predictive performance: training data scale, output resolution, splicing event quantification, and model complexity.
• Examination of how models are applied in variant effect annotation and therapeutic development (e.g., antisense oligonucleotides).

Main Results:

• Computational approaches have advanced from simple heuristics to complex AI models.
• Quantitative assessment of splicing events and increased model complexity are critical for biological interpretability and computational feasibility.
• Splicing prediction models facilitate applications like variant effect annotation and antisense oligonucleotide design.

Conclusions:

• Significant progress has been made in computational splicing prediction, enabling translational applications in genomics and medicine.
• Key challenges remain, including interpreting deep-intronic mutations, isoform-level reconstruction, and integrating multimodal data.
• Future opportunities lie in refining model interpretability, addressing complex genetic variations, and leveraging diverse data types for enhanced prediction accuracy.