Finding the cell free DNA (cfDNA) that most cell-types release into blood plasma is always a challenging task, but certain cancers release so little that it takes looking for a needle in a haystack to the next level. Methods for sorting the “hay” have often come up short, but two new studies successfully developed a new cfDNA “magnet”: a cell-free MeDIP-seq protocol optimized to find cfDNA from even the most challenging tumours.
There has been a huge interest in “liquid biopsies”, extracting cfDNA from plasma and identifying the cancer early by detecting the DNA fragments from dying cells that have entered circulation. DNA methylation differs greatly in most cancers, making it an ideal choice to detect the tumor cfDNA that circulates in blood. A major challenge, however, is the very small amount of cfDNA some tumor types release, which is below the threshold of detection for most DNA methylation assays. Thankfully, two new studies now bring forth a modified MeDIP-seq protocol to overcome this obstacle.
The two new studies were published concurrently from the labs of Daniel De Carvalho, Toni Choueiri, Matthew Freedman and Gelareh Zadeh. The authors built on their previously developed cfMeDIP-seq method, which is a MeDIP-seq protocol that can work from the low level of cfDNA circulating in blood. This technique uses low (1-10ng) input of DNA and, for contrast, MeDIP-seq typically requires 100 ng. This low input is possible due the addition of “filler DNA”, a mixture of in vitro methylated and unmethylated PCR amplicons with size distributions similar to those of cfDNA. This DNA acts as a carrier for the immunoprecipitation reaction, boosting its specificity. They also optimized other aspects of the protocol to improve translation to clinical use. Here, they wanted to validate the utility of cfMeDIP-seq for early detection of two particularly challenging cancers: renal cell carcinomas (RCCs) and of central nervous system (CNS) tumours.
Detection of Renal Cell Carcinoma using cfDNA
35% of RCC cases are not diagnosed until the cells has spread beyond the kidney, and there are no blood or radiologic tests to screen for it. RCC sheds less cfDNA than almost any other tumour. In this paper, the talented team used 148 samples: 99 stage I–IV RCC cases, 21 stage IV urothelial bladder cancer (UBC) samples 28 controls. Here’s the summary of their machine learning approaches:
- Using 80% of the plasma RCC cases and controls as a discovery cohort, the top 300 differentially methylated regions (DMRs) were identified and used to build a classifier that assigned a methylation score to the remaining 20%. This was iterated 100 times with different samples being used for each group.
- 67/69 RCC samples are assigned a higher median methylation score than all control samples, an AUROC of 0.990. AUROC is used to evaluate machine learning models, with 0.5 being random and 1.0 being perfect classification
- The same analysis was performed comparing RCC tumours to UBC (instead of controls), yielding similar discrimination, an AUROC of 0.979
- In urine, comparing RCC and controls yields an AUROC of 0.858
Discrimination of CNS Tumors Using cfDNA
The second paper published concurrently focused on validation of cfMeDIP-seq in CNS cancer, another challenging tumour type. Currently, the diagnosis of subtyping of a brain lesion is identified using invasive neurosurgery. Use of non-invasive methods to determine tumour subtypes would better prepare surgeons for removing the tumour, or even remove the need for surgery in less aggressive cases. Previously, the authors found that cfMeDIP-seq could identify CNS tumour DNA in plasma. Now, they wanted to determine if it could discriminate between CNS tumors, some of which share similar cells-of-origin. They used plasma from 60 patients with diffuse glioma as well as previous data from CNS and other cancers.
- Similar to the above study, the authors identified the top 300 DMRs in 80% of the diffuse gliomas vs. control and other cancers, and used the other 20% as validation, iterated 50 times
- The models generated distinguish diffuse gliomas from controls and non-CNS cancers, with AUROC above 0.95 for all
- To confirm that tumour-derived DNA drives this classification, they restricted their dataset to windows unmethylated in healthy plasma
- These DMRs still show similar levels of discrimination, indicating the signal in the cfDNA is tumour-derived
- To further demonstrate this, they defined a glioma-specific DMR signature
- Finally, they sought to determine if cfMeDIP-seq could be used to further distinguish common primary intracranial tumors with 161 such plasma samples
- Using the same statistical approach, they found AUROC ranging from 0.71-0.95 depending on the tumor type
Together, these studies show that cfMeDIP-seq has promise to accurately diagnose and enable early detection of cancers that release little cfDNA. This represents an important validation of this technology for eventual clinical use. cfMeDIP-seq shows how modification of an existing molecular technique can get around a tricky problem. The researchers show that you don’t have to pick through the hay, or go find a haystack with more needles, just bring a magnet.