Using Machine Learning to Identify True Somatic Variants from Next-Generation Sequencing

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Molecular profiling is crucial for cancer risk stratification and treatment selection.
• Identifying true genetic variants is challenging due to cancer genome complexity and technical artifacts.
• Current manual screening methods in clinical labs are costly, subjective, and not scalable.

Purpose of the Study:

• To develop and evaluate a machine learning-based method for distinguishing artifacts from bona fide single-nucleotide variants (SNVs).
• To improve the quality and efficiency of variant review in clinical settings using next-generation sequencing data.

Main Methods:

• A cohort of 11,278 SNVs from tumor specimens was used for training, validation, and testing.
• A 3-class machine learning model (real, artifact, uncertain) was developed and evaluated.
• Prediction intervals were incorporated to label uncertain variants.

Main Results:

• The classifier achieved 100% specificity and 97% sensitivity on the test set (5,587 SNVs).
• 96.6% of SNVs received definitive labels, exempting them from manual review.
• Zero misclassifications occurred between true positives and artifacts in the test set.

Conclusions:

• A computational classifier was developed to identify variant artifacts from tumor sequencing.
• The proposed framework enhances the quality and efficiency of variant review processes in clinical laboratories.
• This method automates the distinction between real and artifactual variants, streamlining genomic data analysis.