Xenograft supervision and joint methylation-fragment-end encoding define a reusable molecular scoring layer for cell-free tumor DNA
- 1Hefei Cancer Hospital of CAS, Institute of Health and Medical Technology, Hefei Institutes of Physical Science, Chinese Academy of Sciences, Hefei, 230031, P. R
Abstract
Circulating tumor DNA is often scarce in plasma, and regional methylation summaries can obscure informative differences between individual molecules. Here we present ctDNA-BERT, a Transformer that scores individual bisulfite reads without predefined differentially methylated regions by jointly encoding CpG methylation state, nucleotide context and the fragment 3' end. Xenografts provide molecule-level tumor supervision. The resulting frozen scorer supported tumor-fraction estimation in synthetic and patient-derived xenograft mixtures and enriched tumor-concordant methylation and chromatin-accessibility signals in independent hepatocellular carcinoma and stage III pan-cancer plasma without cohort-specific retraining. Features derived from selected reads discriminated stage I cancers from healthy controls in a retrospective cohort of 1,477 specimens, achieving an out-of-fold area under the receiver operating characteristic curve of 0.884. By separating molecular scoring from sample-level inference, ctDNA-BERT provides a reusable read-level scoring layer for cell-free DNA methylation analysis.
Related
- High-performance xenograft read-sorting for DNA methylation analysis
- GRASP: Graph Representation Learning with Assay Supervision for Molecular Properties
- STAMP as a topology-aware molecular language for fragment-level learning and editing
- DNA methylation reveals molecular heterogeneity and urate excretion-associated subtypes of gout
- Regulatory mode and enhancer architecture determine limits of blood DNA methylation as a molecular proxy
BibTeX
@misc{zhang2026xenograft,
title = {{Xenograft supervision and joint methylation-fragment-end encoding define a reusable molecular scoring layer for cell-free tumor DNA}},
author = {X. Zhang and X. Sun and R. Wang and N. Zhao and F. Liu and X. Li and J. Huang and J. Sun and H. Li and S. Lou and H. Wang and H. Zhao and H. Shu and P. Zhen and H. Gu and J. Shi and T. Wang and F. Zhang},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.26.761533},
url = {https://www.biorxiv.org/content/10.64898/2026.10.26.761533v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Xenograft supervision and joint methylation-fragment-end encoding define a reusable molecular scoring layer for cell-free tumor DNA
AU - X. Zhang
AU - X. Sun
AU - R. Wang
AU - N. Zhao
AU - F. Liu
AU - X. Li
AU - J. Huang
AU - J. Sun
AU - H. Li
AU - S. Lou
AU - H. Wang
AU - H. Zhao
AU - H. Shu
AU - P. Zhen
AU - H. Gu
AU - J. Shi
AU - T. Wang
AU - F. Zhang
PY - 2026
DA - 2026/10/26
PB - bioRxiv
DO - 10.64898/2026.10.26.761533
UR - https://www.biorxiv.org/content/10.64898/2026.10.26.761533v1
AB - Circulating tumor DNA is often scarce in plasma, and regional methylation summaries can obscure informative differences between individual molecules. Here we present ctDNA-BERT, a Transformer that scores individual bisulfite reads without predefined differentially methylated regions by jointly encoding CpG methylation state, nucleotide context and the fragment 3' end. Xenografts provide molecule-level tumor supervision. The resulting frozen scorer supported tumor-fraction estimation in synthetic and patient-derived xenograft mixtures and enriched tumor-concordant methylation and chromatin-accessibility signals in independent hepatocellular carcinoma and stage III pan-cancer plasma without cohort-specific retraining. Features derived from selected reads discriminated stage I cancers from healthy controls in a retrospective cohort of 1,477 specimens, achieving an out-of-fold area under the receiver operating characteristic curve of 0.884. By separating molecular scoring from sample-level inference, ctDNA-BERT provides a reusable read-level scoring layer for cell-free DNA methylation analysis.
ER -