Researchers have developed purNPE, a Bayesian deep-learning framework designed to enable uncertainty-aware tumor fraction estimation from whole-genome cell-free DNA sequencing without relying on labeled cancer cell-free DNA samples for training. The computational system addresses key challenges in liquid biopsy, where estimating tumor fraction is critical but frequently hampered by weak biological signals and baseline sequencing noise at low tumor fractions.
Existing computational approaches for estimating tumor fraction often require matched controls or extensive labeled datasets for model training, while also lacking mechanisms for uncertainty quantification. To resolve these operational constraints, purNPE was formulated to operate without requiring labeled cancer training samples or reference sample sets during inference.
The methodology relies on a two-part generative modeling strategy. One component simulates diverse tumor copy-number profiles based on evolutionary genealogies, while a second, data-driven component learns and replicates realistic sequencing background noise patterns derived directly from cancer-free cell-free DNA.
By training a Neural Posterior Estimator on these synthetic tumor profiles augmented with learned noise patterns, purNPE conducts amortized inference in milliseconds. This architecture allows the framework to generate tumor fraction estimates rapidly without depending on a reference sample set at the time of inference.
In validation testing across a real-world pan-cancer cohort, purNPE demonstrated performance comparable to existing analytical methods when evaluated against an orthogonal mutant-allele-fraction proxy, achieving a mean absolute error of 0.066. Furthermore, in semi-synthetic and in silico experiments, the model produced calibrated uncertainty estimates and displayed a consistent dose-response across low tumor-fraction spike-ins, achieving separation of approximately 1% tumor fraction spike-ins from blanks under the evaluated test conditions.
The authors characterized the platform as a proof-of-concept framework demonstrating the application of simulation-based inference to obtain calibrated, uncertainty-aware tumor fraction estimates. The researchers emphasized that evaluation in larger prospective cohorts with stronger orthogonal ground truth will be required before the analytical sensitivity or clinical utility of purNPE can be formally established.
Subject of Research: Uncertainty-aware tumor fraction estimation in cell-free DNA using Bayesian deep learning
Article Title: Simulation-based Bayesian deep learning enables uncertainty-aware tumor fraction estimation in cell-free DNA
Article References: Volkov, H., Raitses-Gurevich, M., Grad, M., Shlayem, R., Danilevsky, A., Rubinek, T., Gorfine, M., & Shomron, N. (2026). Simulation-based Bayesian deep learning enables uncertainty-aware tumor fraction estimation in cell-free DNA. BioData Mining, 19(1), Article 84. https://doi.org/10.1186/s13040-026-00595-5
Image Credits: AI Generated
DOI: 10.1186/s13040-026-00595-5
Keywords: cell-free DNA, liquid biopsy, tumor fraction estimation, Bayesian deep learning, purNPE, simulation-based inference, neural posterior estimation, Simulation-based, Bayesian, deep, learning, enables
News Source: Nathaniel Bowman. (November 7, 2026). Simulation-Based Deep Learning Estimates Cell-Free DNA Tumor Fraction with Uncertainty. Scienmag.
