Rethinking DeepVariant: Efficient Neural Architectures for Intelligent Variant Calling.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• DeepVariant uses a neural network for genetic variant identification.
• Its core Inception V3 architecture has not been updated despite widespread use.
• Training data expansion is the primary method of DeepVariant improvement.

Purpose of the Study:

• To modernize DeepVariant by enabling alternative neural network backbones.
• To evaluate the performance of a new architecture in genetic variant calling.
• To assess improvements in accuracy, efficiency, and robustness.

Main Methods:

• Replaced the Inception V3 model in DeepVariant with an EfficientNet model.
• Evaluated the modernized DeepVariant on the Genome in a Bottle (GIAB) benchmark dataset.
• Compared performance metrics including convergence speed, parameter count, and SNP F1-score.

Main Results:

• The EfficientNet backbone demonstrated faster convergence and a twofold reduction in parameters.
• Improved accuracy in genetic variant identification was observed.
• The updated workflow achieved a +0.1% improvement in SNP F1-score, identifying hundreds of additional true variants per genome.

Conclusions:

• Optimizing the neural network architecture alone significantly enhances DeepVariant's performance.
• Modernized DeepVariant offers improved accuracy, robustness, and efficiency for variant calling.
• This advancement contributes to higher quality sequencing data analysis.