A new machine-learning algorithm that recognizes patterns in genome-wide DNA-methylation data can help pathologists classify brain and spinal cord tumors. The diagnostic tool is freely available online, and could prove valuable for histologically ambiguous cases.

Artificial intelligence can recognize molecular patterns in DNA methylation data to classify tumors of the central nervous system (CNS) with greater diagnostic precision than a human pathologist.

The finding has led to the creation of a free online diagnostic tool that allows pathologists to upload their raw microarray data on the methylation status of hundreds of thousands of single nucleotides in the genome and receive a report classifying a patient's tumor within minutes (Nature 2018;555:469–74; www.molecularneuropathology.org).

This machine-learning method will not supplant standard protocols in neuropathology anytime soon, experts say, but it could be a valuable aid when existing techniques yield inconclusive results. “It's a landmark article,” says David Louis, MD, of Massachusetts General Hospital in Boston, MA, who was not involved in the new study. “It will most likely augment the pathologist's toolbox.”

Next-generation sequencing and epigenetic analyses remain uncommon outside a research setting, but the case for these more advanced molecular tools is growing, say Stefan Pfister, MD, and Andreas von Deimling, MD, from the German Cancer Research Center (DKFZ) and Heidelberg University Hospital, the lead authors of the new study.

To develop their diagnostic tool, the researchers trained a computer algorithm to find methylation patterns in pathologist-classified samples of CNS tumors. The algorithm assigned the tumors to 82 classes, only one third of which exactly matched those defined by the World Health Organization (WHO) classification system. The remainder largely represented subclasses or combinations of established categories, although some were tumor types that the WHO has not previously recognized.

The DKFZ team then pitted their algorithm against a pathologist in 1,104 test cases. For 838 of those, the human and the computer agreed, although the computer sometimes also assigned tumors to a subcategory. In 139 cases, there was a mismatch, and further molecular analyses showed that the computer was correct 93% of the time. In the remaining 127 cases, the algorithm proved inconclusive, often because the samples came from rare pediatric tumors that are not yet part of the reference taxonomy.

Independent groups at five clinics across Europe and the United States validated the tool on another 401 cases and came to similar conclusions about its diagnostic accuracy. The DKFZ team is now prospectively testing the platform in a study called Molecular Neuropathology 2.0, which is enrolling about 80% of children diagnosed with brain tumors in Germany.

Stephen Yip, MD, PhD, of the University of British Columbia in Vancouver, Canada, who wrote an accompanying commentary, expects the tool to gain popularity for thorny cases (Nature 2018;555:446–7). “Simple cases that are immediately recognizable by unique microscopic features, or in which we have simple and cheap adjunct diagnostic tools, will not gain too much from the methylation array at present,” he says. “It's really the histologically ambiguous cases or those that present in an unusual manner that would benefit.”

The methylation-based classification tool is not restricted to CNS tumors. In another paper, the DKFZ team showed that methylation profiling provides valuable diagnostic information for sarcomas, and they have developed a machine-learning algorithm to classify those cancers as well (Mod Pathol 2018 March 23 [Epub ahead of print]).

“This principle of using methylation patterns to determine cell of origin will be useful for any tumor for which there is diagnostic uncertainty,” Pfister says. –Elie Dolgin

For more news on cancer research, visit Cancer Discovery online at http://cancerdiscovery.aacrjournals.org/CDNews.