BayesCNV: A Bayesian Hierarchical Model for Sensitive and Specific Copy Number Estimation in Cell Free DNA

Background/Objectives: Detecting copy number variations (CNVs) from next-generation sequencing (NGS) is challenging, particularly in targeted sequencing panels, especially for cell-free DNA (cfDNA), where the signal is weak and noise is high. Methods: We present BayesCNV, a Bayesian hierarchical model for gene-level copy ratio estimation from targeted amplicon read depths compared to a CNV-neutral reference sample. The model provides posterior uncertainty for each gene and supports interpretable calling based on effect size and posterior confidence. The model also provides a principled quality-control strategy based on the marginal log likelihood of each sample, with low values indicating low confidence in the calls. BayesCNV uses thermodynamic integration, a technique to reliably estimate this quantity. We benchmark our method against two publicly available CNV callers using Seracare^® reference samples with known CNVs on the OncoReveal^® Core Lbx panel. Results: Our method achieves a sensitivity of 0.87 and specificity of 0.996, dramatically outperforming two competitor methods, IonCopy and DeviCNV. In a separate FFPE dataset using the OncoReveal^® Essential Lbx panel, we show that the marginal log likelihood cleanly separates, degraded from high-quality samples, even when conventional sequencing QC metrics do not. Conclusions: BayesCNV provides accurate and interpretable gene-level CNV estimates and uncertainty quantification, along with an evidence-based quality control metric that improves robustness in targeted cfDNA workflows.