LOS ANGELES – Researchers from the Ontario Institute for Cancer Research in Canada and their collaborators have identified plasma cell-free DNA (cfDNA) methylation signatures that can be predictive of certain cancers years before their clinical diagnosis.
Presenting the results at the American Society of Human Genetics annual meeting on Wednesday, Nicholas Cheng, a graduate student in Philip Awadalla’s lab at the Ontario Institute for Cancer Research, said that the researchers also developed a machine learning-based model that could potentially help predict cancers in individuals prior to diagnosis.
“A lot of early cancer detection studies have primarily focused on blood samples collected at the time of diagnosis when someone has already developed cancer,” said Cheng, who is also the first author of a related preprint currently available on Research Square. “If we really want to develop these biomarkers for early cancer detection, we need to assess how well these markers work in asymptomatic individuals before they even develop cancer.”
For its study, Cheng’s team tapped into samples from the Canadian Partnership for Tomorrow’s Health Project (CanPaTH), a large, publicly available population study cohort that contains over 40,000 blood samples from individuals who were cancer-free when they enrolled in the study.
From there, the researchers identified participants who subsequently developed breast, prostate, or pancreatic cancer and performed cell-free methylated DNA immunoprecipitation and high-throughput sequencing (cfMeDIP-seq), a highly sensitive assay previously developed by Cheng’s collaborators for profiling the cfDNA methylome, on over 400 blood plasma samples collected up to seven years prior to the cancer diagnosis. In addition, the team selected cancer-free control samples from CanPaTH that were matched for age, sex, and sample collection time, as well as other health and lifestyle factors.
According to Cheng, the first step for the team was to figure out whether there were any methylation signatures in the cfDNA that could discriminate between the cancer-free controls and the pre-diagnosis blood samples.
Using differential methylation analysis, the researchers identified the top 200 hypermethylated regions for each cancer and found that these signatures could help distinguish cancer-free controls from pre-diagnosis cancer cases more than five years prior to diagnosis.
The scientists then investigated whether the captured hypermethylated cfDNA regions can be observed in bulk cancer tissue samples. Since the researchers did not have access to the matching tumor samples of the CanPaTH participants, they leveraged publicly accessible cancer tissue samples as well as adjacent normal tissues and peripheral blood leukocytes from the Cancer Genome Atlas (TCGA).
The analysis revealed a “significant overlap” among the top 100 hypermethylated regions in the pre-diagnosis blood samples with the cancer bulk tissues, Cheng said. In addition, the team observed that some of these DNA methylation patterns were also highly associated with gene expression changes, pointing to early genomic processes that may contribute to cancer development.
Encouraged by these results, the researchers used the methylation signatures to develop a machine learning-based predictive model and investigated if it could accurately predict an individual’s cancer based on their pre-diagnosis blood cfDNA methylome.
The team tested the model in a breast cancer discovery cohort that consisted of 67 pre-diagnosis samples and 59 non-cancer controls. Overall, the results showed that the model performed “relatively well,” Cheng said, with an area under the receiver operating characteristic curve (AUC) of around 0.724, over 25 percent sensitivity, and over 95 percent specificity.
More importantly, Cheng said, when looking into specific patient subgroups, the team noticed that the model delivered better results in predicting cancers in individuals under the age of 50. “This is particularly interesting because, in Canada, the recommended age for mammogram screening for women is over the age of 50,” he pointed out. “So, we were able to detect some of these cancers in individuals before mammogram screening.”
Furthermore, Cheng said the model was shown to help predict cancers in individuals who had a negative mammogram screen within one year of the blood collection, indicating that the assay could potentially help detect cancers that would have been missed by mammography.
The researchers further validated the model in an independent cohort consisting of 19 pre-diagnosis breast cancer samples and 15 controls, where the test achieved 50 percent sensitivity and perfect specificity, Cheng noted.
Even though the current work focuses on breast, prostate, and pancreatic cancer cases, moving forward, the team is planning to extend the assay to pan-cancer applications, he added.