Fragmentomic-based algorithm to computationally predict tumor-somatic, germline, and clonal hematopoiesis variant...

Area of Science:

• Oncology
• Genomics
• Bioinformatics

Background:

• Liquid biopsy (LBx) offers a non-invasive method for tumor genomic profiling, serving as an alternative to tissue biopsy.
• Clonal hematopoiesis (CH) variants can be detected in LBx, potentially confounding tumor-specific variant identification and impacting treatment decisions.

Purpose of the Study:

• To develop and validate a machine learning algorithm (Variant Origin Prediction - VOP) to accurately distinguish tumor-somatic variants from CH variants in LBx.
• To enhance the reliability of LBx for cancer treatment decision-making by clarifying variant origins.

Main Methods:

• Sequenced paired plasma and white blood cell (WBC) DNA to train and validate the VOP algorithm.
• Utilized fragmentomics data within the VOP algorithm to predict the origin of short variants (SVs) detected by LBx.
• Validated VOP accuracy using paired WBC DNA and assessed reproducibility with LBx replicates.

Main Results:

• The VOP algorithm demonstrated high sensitivity and positive predictive value (>93% and >91%, respectively) in differentiating tumor and CH variants.
• VOP accurately identified variants with low variant allele frequencies (VAFs ≤1%) and in genes like TP53, which can harbor both CH and tumor-somatic variants.
• In a longitudinal study of metastatic castration-resistant prostate cancer (mCRPC), VOP accurately predicted variant origins and enabled separate tracking of tumor and CH variants over time.

Conclusions:

• Variant Origin Prediction (VOP) provides a highly accurate and reproducible method for determining the origin of SVs in LBx, without requiring concurrent WBC sequencing.
• VOP can minimize the inappropriate use of targeted therapies and associated toxicities in patients with CH variants.
• The algorithm facilitates precise tumor profiling and monitoring, improving patient management.