There are about 100 different known tumor types of the central nervous system. Based on tissue characteristics or gene mutations, these tumor types can show widely varying responses to radiotherapy and chemotherapy. Due to this considerable variability, standardization of the diagnostic process has been particularly challenging.
New findings now propose a new machine-learning-based approach for comprehensively classifying CNS tumors that is applicable to routine cancer diagnostics. The recently published work in Nature comes from a large collaboration led by researchers from the Hopp Children's Cancer Center at the NCT Heidelberg in Germany.
"We hope that our new molecular classification method will help improve diagnostic accuracy in CNS tumors and, thus, also improve the chances for successful treatment," said senior co-author Andreas von Deimling.
Using a computer-based method, the team combed for DNA methylation patterns to draw conclusions on tumor cellular origins. The algorithms they developed were able to reliably differentiate 82 types of CNS tumors based on their methylation patterns.
"Particularly in tumors which we cannot easily assign to a diagnostic category based solely on microscopic examination, methylation analysis is often helpful to make a precise diagnosis,” explained lead co-author David Capper. “The analysis of approximately 2,800 reference tumor samples additionally made it possible to classify tumors into specific subgroups that are not yet included in the classifications that have been used so far."
To test the diagnostic precision of their process, the team looked at the DNA methylation patterns of an additional 1,100 tumor samples. As a result, they were able to correctly change the initial diagnosis in up to 12% of the cases. Further molecular-based diagnostic tests also showed that the methylation-based classification characterized the tumors even better than the initial diagnosis based on microscopy.
The team believes that their findings can provide a blueprint for creating algorithm-based tumor classifiers across other cancer entities. "We are convinced that our new method is well suited to be used in the clinic. We have made our classification system available online in order to enable researchers to analyze their data at our platform, " said senior co-author Stefan Pfister.