Introduction: The acquisition of somatic mutations in hematopoietic stem and progenitor stem cells with resultant clonal expansion is known as clonal hematopoiesis (CH). CH is linked to a higher risk of hematologic malignancies and other adverse health outcomes. The prevalence of CH is dependent on sequencing technique. With ultra-high-depth sequencing, CH can be detected in most adults. However, CH variant calling at very low variant allele fractions (VAF) is challenging due to difficulty in distinguishing low frequency CH mutations from sequencing artifacts. Here we develop and validate a novel CH variant calling approach that combines ensemble based variant calling with advanced artifact filtering and outperforms commonly used somatic variant callers. Methods: We sequenced a tumor-normal dilution series from six AML patients and 27 technical controls with a targeted panel utilizing unique molecular indexes that included nine common CH genes (DNMT3A, TET2, ASXL1, TP53, PPM1D, JAK2, SF3B1, SRSF2) with a mean unique coverage of 18,268x. In total we identified 32 mutations in these six samples validated by an independent sequencing panel performed for the purpose of clinical testing. The variant allele fraction (VAF) in the resulting tumor-dilution samples ranged from 0.1%-50% VAF. Results: For higher VAF mutations (>5%), Mutect2, Lofreq, and Vardict showed reasonable sensitivity (range: 0.70-1.00) and positive predictive values (PPV) (range: 0.80-1.00). However, at lower VAF, PPV was poor; for mutations between 0.1-0.4% VAF, the PPV ranged between 0.01-0.22. We applied additional filters based on sequencing quality, depth, regional complexity, and an empiric estimate of sequencing error at the position of a given called variant using technical controls. Taking the consensus of three callers and including our additional filters, we not only retained sensitivity and PPV for higher VAF variants above 1% (0.92 sensitivity, 0.95 PPV) but improved the PPV with retained sensitivity for variants between 1-0.1% VAF (0.91 sensitivity, 0.83 PPV). Applying our consensus approach to blood samples drawn from 31 healthy individuals sequenced using the same targeted panel with variants validated using an orthogonal approach, we saw excellent sensitivity and PPV across a wide range of VAFs (Sensitivity 0.99 [>1% VAF], 0.92 [0.2-1% VAF], and 0.82 [0.1-0.2% VAF]; PPV 1.00 [>1% VAF], 0.98 [0.2-1% VAF], 0.97 [0.2-0.1% VAF]). We compared our consensus calling and advanced filtering approach to a machine learning model. A decision-tree-based ensemble machine learning algorithm, XGBoost, was trained on the tumor-dilution samples with validation using 31 blood samples with orthogonal sequencing. The variant caller output (allele counts, flags, etc.), additional false positive filtering metrics, and statistical results from the technical controls were used as features. Sensitivity and PPV was high for mutations above 1% VAF (0.97 sensitivity and 0.82 PPV) with the performance declining for lower VAF mutations (0.72 sensitivity and 0.65 PPV). Thus, a machine learning approach did not outperform consensus calling with advanced artifact filtering. Finally, we developed a custom annotation pipeline for CH variant pathogenicity classification incorporating previously reported CH mutations in large cohorts, cancer somatic mutation frequency databases and curated cancer driver knowledge databases. Conclusion: We developed ArCCH, an advanced artifact filtering and consensus CH calling pipeline that includes custom annotation features to facilitate CH pathogenicity determination. Our results show that our consensus approach with advanced error correction filters substantially improves the performance of CH variant calling beyond commonly used single variant callers. This end-to-end flexible workflow package is publicly available on Terra and will facilitate future CH studies.
Read full abstract
Feb 1, 2023
Ivo Colombo + 1
Open Access