Abstract
Abstract
Tumor phenotype and therapeutic vulnerability are shaped by dynamic transcriptional states that are not captured by genetic profiling alone and are difficult to monitor longitudinally. Existing plasma-based approaches have largely been developed using samples with high circulating tumor DNA fractions, evaluated across expression quantiles rather than individual genes, and rarely benchmarked head-to-head. Using timepoint-matched plasma epigenomes and tumor transcriptomes, we identify biologically grounded relationships between circulating chromatin, fragmentation, and transcription. H3K4me3 and H3K36me3 encode distinct positional information reflecting promoter activity and transcriptional elongation, while gene-body fragment periodicity predicts expression. We integrate these relationships into APEX, a machine-learning framework for genome-wide single-gene expression inference. Across 559 plasma samples spanning 24 cancer types and clinically relevant tumor fractions, APEX outperforms existing approaches for individual-gene inference, resolves pancreatic cancer transcriptional subtypes, tracks therapeutic targets in small-cell lung cancer, and identifies NECTIN4 expression as a biomarker of response to enfortumab vedotin in urothelial cancer.