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

Area of Science:

• Genomics
• Computational Biology
• Molecular Genetics

Background:

• Alternative splicing is vital for protein diversity and gene regulation in eukaryotes.
• Mutations affecting splicing are linked to numerous genetic disorders.
• Current computational methods for predicting splice site usage accuracy are limited.

Purpose of the Study:

• To develop an accurate deep neural network model, DeltaSplice, for predicting the quantitative impact of mutations on alternative splicing.
• To improve the identification of disease-causing mutations and potential drug targets through enhanced splicing prediction.
• To leverage comparative analysis of homologous genes for reference-informed prediction.

Main Methods:

• Developed DeltaSplice, a deep neural network model utilizing comparative analysis of homologous genes.
• Incorporated a "reference-informed prediction" strategy using known splice site usage.
• Benchmarked DeltaSplice against state-of-the-art methods on diverse prediction tasks.

Main Results:

• DeltaSplice consistently outperformed existing methods in predicting splicing-altering mutations.
• Identified approximately 15% of human brain splicing quantitative trait loci (sQTLs) as causal splicing-altering variants.
• Predicted splicing-altering de novo mutations in genes associated with autism and neurodevelopmental disorders (NDDs), identifying 19 recurrently mutated genes and 8 novel NDD-risk genes.

Conclusions:

• DeltaSplice significantly advances the accuracy of in silico splicing models.
• The model shows potential for improving genetic diagnosis and developing precision medicine approaches for splicing-related disorders.
• This work expands the utility of computational models in understanding gene regulation and disease mechanisms.