Genome-wide analysis of 5-hydroxymethylcytosine signals in early cancer detection through direct analysis of cell-free DNA - 11th International Symposium on Circulating Nucleic Acids in Plasma and Serum (CNAPS)

Anna Bergamaschi ¹ Chin-Jen Ku ² Yuhong Ning ² Francois Collin ² Chris Ellison ¹ Tierney Phillips ¹ Erin McCarthy ¹ Wendy Wang ¹ Michael Antoine ¹ Aaron Scott ² Paul Lloyd ² Gulfem Guler ² Alan Ashworth ³ Stephen Quake ⁴ Samuel Levy ¹

¹Research, Bluestar Genomics, San Diego, CA, USA
²Research, Bluestar Genomics, San Francisco, CA, USA
³Helen Diller Family Comprehensive Cancer Center, UCSF, San Francisco, CA, USA
⁴Departments of Bioengineering and Applied Physics, Stanford University, Palo Alto, CA, USA

Strategies which interrogate dynamic and temporal genome-wide signals in a liquid biopsy context may address precision medicine’s needs to improve cancer care. Methylation and hydroxymethylation of cytosines enable the epigenomic regulation of gene suppression and activation. 5-hydroxymethyl-cytosine (5hmC) is globally decreased in tumor tissue and changes upon tumorigenesis and has been employed as a cancer biomarker. We sought to identify genome-wide 5hmC changes in plasma based cfDNA from cancer patients representing multiple disease types, stages and clinical characteristics in comparison with non-cancer patients. cfDNA was isolated from plasma, enriched for the 5hmC fraction using chemical labelling, sequenced, and aligned to the genome to determine 5hmC counts per genomic feature. Logistic regression methods were applied to classify cancer and control samples with measurably high performance. Predictive performance was established using 5-fold cross-validation coupled with out-of-fold area under the ROC curve (AUROC). 270 controls and 190 cancers across three disease types (breast, colorectal, and lung) were included, where more than 60% of patients had early stage disease (I or II). Predictive performance was established for breast, lung and colorectal cancer to be (AUROC) 0.89 (CI 0.72-0.92), 0.81 (CI 0.73-0.84) and 0.78 (CI 0.66-0.83), respectively. The genes defining each of these predictive models were enriched for pathways relevant to disease specific etiology. The breast cancer cohort consisted primarily of stage I and II patients, including tumors < 2cm that exhibited a high cancer probability score. Further, a logistic regression to enable the prediction of control or cancer of origin was able to classify any one sample with an AUROC = 0.7-0.8. Misclassification of samples was primarily in the control group. These findings suggest that 5hmC changes in cfDNA enable classification of early stages of breast, colorectal and lung cancer. Further 5hmC data coupled with machine learning provides powerful classification schemes for pan-cancer detection.