Control-independent mosaic single nucleotide variant detection with DeepMosaic

Area of Science:

• Genetics
• Computational Biology
• Bioinformatics

Background:

• Mosaic variants (MVs) arise from mutagenic processes during development and aging.
• Detecting non-clonally expanded MVs in noncancerous conditions is computationally difficult.
• MVs are implicated in diseases like cancer and autism.

Purpose of the Study:

• To develop an accurate computational method for detecting mosaic variants in noncancer samples.
• To improve upon existing methods for MV detection, particularly for sparse, nonclonally expanded variants.

Main Methods:

• DeepMosaic integrates an image-based visualization module for single nucleotide MVs.
• A convolutional neural network classifies MVs for control-independent detection.
• The model was trained on simulated and experimental MVs and benchmarked on extensive genomic and exomic datasets.

Main Results:

• DeepMosaic demonstrated superior accuracy compared to existing methods on biological data.
• Achieved high sensitivity (0.78), specificity (0.83), and positive predictive value (0.96) on noncancer whole-genome sequencing data.
• Doubled the validation rate on noncancer whole-exome sequencing data compared to previous best-practice methods.

Conclusions:

• DeepMosaic provides an accurate and effective MV classifier for noncancerous samples.
• It can serve as a valuable alternative or supplement to current MV detection techniques.
• Facilitates better understanding of mutagenic processes and MV-associated diseases.