Using machine learning to facilitate variant classification in pediatric cancers

Molecular profiling has become essential for tumor risk stratification and treatment selection. Currently, clinical laboratories rely on manual screening of the variant, which is costly, tedious, and not scalable. We developed a machine learning-based classifier using 11,278 clinically reviewed variants to automate somatic variant screening in cancer diagnostics.




The model learned characteristics separating true mutations from sequencing artifacts to automatically identify real single nucleotide variants from NGS. The optimized three-class model (real, artifacts, uncertain) demonstrated high accuracy and robustness. Implementation of this approach in clinical labs could improve the overall quality and efficiency. Read more