NEW YORK – A team of researchers from the University of Texas MD Anderson Cancer Center has developed a method to classify four subtypes of small cell lung cancer (SCLC) using DNA methylation markers and predictive machine learning models.
In a paper published last month in Cancer Cell, the researchers describe a method to create a diagnostic assay "that can recapitulate this transcriptionally defined phenotype, which is obviously much more challenging than just looking at mutation," said Simon Heeke, an assistant professor at MD Anderson and one of the developers of the method.
In previous research, the team had used mRNA profiling to show that SCLC consists of four transcriptional subtypes, but it needed to translate that information into a diagnostic test that could be applied clinically.
Small cell lung cancer can be challenging to diagnose because it is usually detected when it has already advanced to a metastatic stage, Heeke said, adding diagnosis is limited by the amount of tissue from biopsy specimens that is available. On the plus side, however, the cancer is very distinct and likely has the highest shedding of circulating tumor DNA among all solid tumors.
The use of mRNA expression signatures and multi-marker immunohistochemistry has been considered for subtyping SCLC via tissue, but the approaches "have shortcomings limiting their routine clinical adoption, such as mRNA degradation commonly seen in preserved SCLC specimens and the use of subjective and time-intensive scoring methods used for multi-marker IHC assays," the researchers wrote in last month's paper.
Liquid biopsy has also been explored extensively for SCLC subtyping, the researchers noted, but work by other research groups was "limited by the absence of tumor specimens for direct comparison or profound subtyping of patients based on clinically validated gene expression-based subtyping, which hampers the routine implementation of SCLC subtyping," they wrote.
As a result, the MD Anderson researchers decided to use DNA methylation from both tumor and circulating tumor DNA for its method.
To develop their method, the researchers first engaged in "consensus classification," Heeke said. Because SCLC subtypes "can be quite heterogeneous," it's not always clear which subtype a patient sample may fall into, he added. The team trained 500 models for each subtype to provide a yes or no answer on whether the sample fell into that subtype, and a patient sample was only confirmed for a specific subtype if more than 50 percent of the models agreed. "This left us with the opportunity to say, 'We don't know which subtype this is,'" Heeke said.
The consensus classification step was performed on a cohort of 179 patients that had RNA and DNA extracted from tissue samples, as well as DNA extracted from plasma, and the team used its predictive model to classify the cohort based on RNA gene expression data, which "works phenomenally," according to Heeke.
However, he noted that he and his colleagues were concerned about the robustness of their method because RNA is heavily degraded in clinical tissue specimens. As a result, they decided to move forward utilizing DNA methylation because "DNA methylation is one of those prime mechanisms that we have to guide gene transcription" and can serve as "a surrogate for the transcriptional phenotypes" the team was looking at, Heeke said. DNA is also well preserved in tissue specimens, making it easier to use for clinical testing. Once the researchers confirmed their method could be used with DNA methylation in tissue specimens, they were able to use the method for DNA methylation in plasma to make a liquid biopsy version of its classifier.
The team has currently developed three classifiers — the SCLC-GRC, for RNA-based tissue classification; the SCLC-DMC, for DNA-based tissue classification; and the SCLC-cfDMC, for DNA-based plasma classification. Heeke noted that the researchers "still have some work to do" to get any of the classifiers to clinical application and regulatory approval. The researchers noted in the Cancer Cell paper that the SCLC-DMC was able to classify tumor samples that failed classification using RNA, "suggesting potential advantages of DNA methylation over gene expression signatures."
For the liquid biopsy test, the researchers extracted circulating tumor DNA from a patient's plasma sample and used reduced representation bisulfite sequencing to get the data for the DNA methylation. "It's not completely whole-genome DNA methylation analysis, but it's enriching for … sites with high DNA methylation content," Heeke noted.
They then fed the methylation information into the classifier that determines which subtype is indicated based on the methylation markers. If more than 50 percent of the models agreed on a certain subtype, they assigned that subtype to the patient, Heeke said.
Heeke said that the researchers are hopeful that their classifiers can be used to stratify patients for pharmaceutical clinical trials and personalized therapies, particularly as many other research teams have used models to define potential therapies that may work for certain subtypes. The technology described in last month's Cancer Cell paper is exclusively for patients that already have an SCLC diagnosis, but Heeke noted that the researchers have seen "that small cell lung cancer has a very distinct DNA methylation profile" and that it could be used to distinguish between small cell lung cancer and non-small cell lung cancer in the future, which the team is working on.
The researchers are currently working to get a CLIA-validated version of the assay that can be further validated in prospective clinical trials and implemented as a diagnostic tool for personalizing therapy, while they continue to validate their method with additional data sets to further confirm its reproducibility and have a "good network of partners" who can help translate the method into a technology that can be applied clinically, Heeke said.