Predicting somatic mutation origins in cell-free DNA by semi-supervised GAN models

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Distinguishing pathogenic cancer mutations from clonal hematopoiesis (CH) variants in cell-free DNA (cfDNA) is crucial for accurate liquid biopsy diagnoses.
• Misclassification can lead to incorrect diagnoses and suboptimal therapeutic strategies.

Purpose of the Study:

• To develop a machine learning technique for differentiating tumor-derived mutations from CH-related mutations in cfDNA.
• To enhance the accuracy and reliability of liquid biopsy analyses.

Main Methods:

• Developed a deep learning model based on the semi-supervised generative adversarial network (SSGAN) architecture.
• Created an in-house reference catalog of approximately 25,000 single nucleotide variants (SNVs) with known tumor or CH origins.
• Trained the model using genomic coordinates and nucleotide composition of cfDNA variants.

Main Results:

• The SSGAN model achieved a 95% area under the curve (AUC) in classifying unknown cfDNA variants.
• Demonstrated the model's capability to accurately differentiate between tumor and CH mutations.
• Validated the potential of genomic feature prediction for liquid biopsy.

Conclusions:

• Genomic feature prediction using cfDNA data offers a robust alternative to traditional multi-analyte sequencing.
• Advanced data analysis and machine learning hold significant potential for improving genomics and personalized medicine.
• The developed method enhances the accuracy of distinguishing CH from tumor mutations in liquid biopsy data.