Reference-informed prediction of alternative splicing and splicing-altering mutations from sequences.

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Alternative splicing is vital for protein diversity and gene regulation in eukaryotes.
• Mutations affecting splicing cause genetic diseases, but accurate prediction remains challenging.
• Current computational methods for predicting splice site usage have limited accuracy.

Approach:

• Developed DeltaSplice, a deep neural network model for quantitative prediction of alternative splicing.
• DeltaSplice uses a "reference-informed prediction" approach, leveraging known splice site usage from homologous genes.
• The model learns mutation impacts on splicing by analyzing comparative genomics of homologous sequences.

Key Points:

• DeltaSplice consistently outperformed state-of-the-art methods in predicting splicing alterations.
• It identified ~15% of brain splicing quantitative trait loci (sQTLs) as causal variants.
• Predicted splicing-altering mutations outside splice sites in neurodevelopmental disorders, identifying new candidate disease genes like MFN1.

Conclusions:

• DeltaSplice enhances the predictive power of in silico splicing models.
• The model shows potential for improving genetic diagnosis and developing splicing-based precision medicine.
• Accurate splicing prediction can accelerate the discovery of disease mechanisms and therapeutic targets.