DNA methylation signature predicts cancer response to demethylation agents from profiling diverse cancer cell lines

Dear Editor, Abnormal DNA methylation, a process whereby tumor suppressors tend to be hypermethylated and silenced, is a hallmark of cancer cells [1]. Removing the methylation by demethylation agents such as azacitidine and decitabine is one of the strategies to treat cancers and has been successfully used to treat certain hematological and solid tumors [2]. However, there are no established DNA methylation markers or signatures that can accurately predict patients’ response to treatment [2]; thus, the identification of reliable predictive biomarkers for effective therapy remains a critical need in clinical practice. Using genome-wide DNA methylation and response data to four demethylation agents (azacitidine, decitabine, RG108 and zebularine) in nearly 600 cancer cell lines, we systematically profiled the response patterns of the demethylation agents, conducted genome-wide association analysis of DNA methylation with the response for each drug, and identified key responsible pathways that could be associated with treatment response. Further, we applied machine learning techniques to develop a model to predict cancer’s response to decitabine (Supplementary Materials and Methods, Supplementary Figure S1). Our results showed that decitabine was the most potent drug among the 4 demethylation agents (smaller area under the curve [AUC] values represent greater drug potency), and hematopoietic/lymphatic cancer cells were the most responsive to all the drugs (decitabine and hematopoietic/lymphatic cancer cells had the smallest median or mean AUC; cancer cell line types were ordered by mean AUC ascendingly from left to right in Supplementary Figure S2A). These findings are consistent with clinical observations suggesting that cell lines could be rep-