Xenograft supervision and joint methylation-fragment-end encoding define a reusable molecular scoring layer for cell-free tumor DNA

ctDNA-BERT scores individual bisulfite reads by jointly encoding CpG methylation state, nucleotide context, and fragment 3' ends, enabling stage I cancer detection with an out-of-fold AUC of 0.884.

• Why it matters: Circulating tumor DNA is often scarce, and regional methylation summaries can obscure informative differences between individual molecules, limiting read-level analysis.
• What they did: The Transformer uses xenografts for molecule-level tumor supervision and a frozen scorer to estimate tumor fraction and enrich tumor-concordant methylation and chromatin-accessibility signals in xenograft mixtures and independent plasma cohorts.
• The result: In a retrospective cohort of 1,477 specimens, read-derived features discriminated stage I cancers from healthy controls with an out-of-fold ROC AUC of 0.884, supporting a reusable read-level scoring layer for cfDNA methylation analysis.