Pathogenicity prediction for noncanonical splice-altering variants based on multimodal feature fusion

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Molecular Genetics

Background:

• Splice-altering variants (SAVs) are a major class of pathogenic genetic variants linked to numerous diseases.
• Current computational tools struggle to accurately assess the pathogenicity of noncanonical SAVs, particularly those outside canonical splice sites.
• This limitation hinders accurate genetic diagnostics and understanding of disease mechanisms.

Purpose of the Study:

• To develop a deep learning framework, MOSAIC, for precise pathogenicity prediction of noncanonical SAVs.
• To improve the assessment of genetic variants impacting RNA splicing.
• To provide a robust computational tool for genetic diagnostics and precision medicine.

Main Methods:

• Developed MOSAIC, a deep learning framework integrating multimodal features.
• Utilized a pretrained DNA language model for long-range contextual signals.
• Employed multi-scale convolutional neural networks for local sequence features and functional annotations, with a transformer encoder and gated fusion module for adaptive integration.

Main Results:

• MOSAIC demonstrated superior performance over existing methods like CADD and SpliceAI across multiple independent datasets.
• The model maintained high accuracy and robustness on rare variants, gene-independent contexts, and challenging datasets.
• Feature importance analysis highlighted the critical role of long-range DNA dependencies and transformer-based integration.

Conclusions:

• MOSAIC provides an accurate and interpretable framework for predicting noncanonical SAV pathogenicity.
• The tool can identify key regulatory motifs, offering mechanistic insights into splicing disruption.
• MOSAIC serves as a dependable computational resource for genetic diagnostics and precision medicine.