High-throughput deep learning variant effect prediction with Sequence UNET.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Understanding genetic mutations is crucial for biology and medicine.
• The vastness of mutation data challenges experimental characterization.
• Existing prediction tools are often computationally intensive and difficult to scale.

Purpose of the Study:

• To introduce Sequence UNET, a scalable deep learning architecture for mutation analysis.
• To classify and predict variant frequency directly from DNA sequences.
• To enable efficient analysis of large-scale variant data.

Main Methods:

• Developed a fully convolutional compression/expansion architecture (Sequence UNET).
• Utilized multi-scale representations for sequence analysis.
• Applied the model to predict variant frequency and pathogenicity.

Main Results:

• Sequence UNET achieves comparable pathogenicity prediction to existing methods.
• Demonstrated scalability by analyzing 8.3 billion variants across 904,134 proteins.
• The model runs efficiently on modest hardware.

Conclusions:

• Sequence UNET offers a highly scalable and efficient approach to analyzing genetic variants.
• The tool facilitates large-scale genomic studies using deep learning.
• It provides a practical solution for mutation prediction in biology and medicine.