Extending DeepTrio for sensitive detection of complex de novo mutation patterns

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• De novo mutations (DNMs) and clustered DNMs (cDNMs) are implicated in genetic disorders.
• Accurate identification of DNMs in whole genome sequencing (WGS) data presents specificity challenges.

Purpose of the Study:

• To develop and evaluate a deep learning framework for enhanced DNM and cDNM detection in WGS data.
• To leverage Google's DeepTrio software for improved variant calling accuracy.

Main Methods:

• A deep learning model was trained for DNM and cDNM detection using WGS trio data from HiSeq and NovaSeq platforms.
• The framework incorporates surrounding sequence information (110 bp up- and downstream) for variant context.

Main Results:

• The DNM detection model achieved 95.7% sensitivity and 89.6% precision.
• An extended model improved cDNM isolation (76.9% precision) with a slight trade-off in overall DNM detection.
• Confidence probabilities allow for user-defined sensitivity-specificity trade-offs.

Conclusions:

• The retrained DeepTrio framework effectively identifies complex mutational signatures with minimal modification.
• This approach offers a promising tool for precise genetic disorder variant analysis.