Abstract 668: Systematic and robust prediction of gene expression using comprehensive epigenomic profiling of plasma ctDNA

Abstract Background: Accurate prediction of RNA transcript levels from circulating tumor DNA (ctDNA) offers a promising approach for dynamic noninvasive cancer biomarker characterization and longitudinal monitoring. By leveraging epigenomic signals proximal to individual genes, we developed transcript level machine learning prediction models for breast and prostate cancer using simulated plasma and validated them in patient plasma. Methods and Results: In breast cancer, our models predicted a panel of 760 key breast cancer oncogenes and transcriptional drivers at simulated ctDNA fractions that capture approximately 45% of stage IV breast cancers profiled, enabling clinically relevant prediction of cancer gene transcription. These models are tuned to separate healthy plasma signal from cancer cells’ contribution to the plasma. The panel is enriched for genes important for transcriptional regulatory identity and breast cancer subtypes, as well as targets for antibody drug conjugates (ADCs) and radio immune conjugates (RICs), including HER2, PSMA , Nectin-4, B7-H3, B7-H4, MET, and EGFR. Our models demonstrated direct applicability in patient plasma with orthogonal validation. For example, they accurately predicted prostate cancer PSA and PSMA PET SUVmean, as well as breast cancer HER2, ER, and PR status. Additionally in a cohort of patients with matched tissue RNA-seq there was a strong correlation between RNA-seq and plasma-based predictions. Further improvements were achieved with bespoke feature engineering and multi-gene aggregation. A PSMA predictor refined with locus-specific features improved performance at lower ctDNA fractions for predicting PSMA PET SUVmean. Similarly, a multi-gene DLL3 predictor in lung cancer reduced the ctDNA threshold from 4% to 1.9% by incorporating predicted expression of three highly correlated genes. In predicting ER+ IHC breast cancer, ESR1 transcript prediction alone performed well at 1% ctDNA (AUC=.78) but was significantly improved (AUC=.90) by incorporating predictions of transcription partners FOXA1 and GATA3. Conclusions: Comprehensive epigenomic profiling enables accurate prediction of tumor-specific RNA transcription in cfDNA, even at low ctDNA levels. This approach enhances ADC target profiling and enables tumor transcriptional profiling in patients from blood. Expanding beyond proof-of-concept models promises improved predictive accuracy at lower tumor fractions and broader applications in diagnostics and therapeutic targeting. Citation Format: Khoi Nguyen, Justin Finkle, Anthony D'Ippolito, Jonathan Beagan, Aparna Gorthi, Travis A. Clark, Amy Donahue, Baovy Tran, Tyrone Tamakloe, Charlene O'Brien, Michael Coyne, Jenna I. Wurster, Kristian Cibulskis, Corrie A. Painter, Mike Zhong, Humphrey Gardner, Jacob Berchuck, Jamey Guess, Jayne Stommel, Gordon B. Mills, SMMART Clinical Trials Program, J. Carl Barrett, Matthew L. Eaton. Systematic and robust prediction of gene expression using comprehensive epigenomic profiling of plasma ctDNA [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 668.