A Transformers-based framework for refinement of genetic variants.

Area of Science:

• Genomics
• Bioinformatics
• Artificial Intelligence

Background:

• Accurate genetic variant calling is essential for distinguishing true variants from sequencing artifacts.
• Traditional methods struggle with accuracy, particularly in low-coverage regions, leading to errors.
• Manual review is laborious, and heuristic filtering lacks optimal solutions for complex data.

Purpose of the Study:

• To develop an automated, accurate, and efficient framework for genetic variant refinement.
• To leverage deep learning, specifically Transformer architectures, for improving variant quality.
• To create a flexible tool that integrates seamlessly into existing genomic analysis pipelines.

Main Methods:

• Developed a Transformer-based framework (VariantTransformer) for genetic variant refinement.
• The model directly processes VCF files, utilizing self-attention mechanisms.
• Trained the framework on 2 million variants from the GIAB HG003 sample (v4.2.1).

Main Results:

• Achieved 89.26% accuracy and a ROC AUC of 0.88 on the training dataset.
• VariantTransformer improved baseline filtering accuracy by 4%-10% across tested samples.
• Outperformed traditional heuristic filters and approached DeepVariant's accuracy when refining existing caller outputs.

Conclusions:

• VariantTransformer offers a flexible and generalizable framework for variant refinement in genomics.
• Demonstrates the potential of Transformer architectures for genomic quality control tasks.
• Provides a blueprint for adapting Transformers to various genomic filtering challenges.