Skip to main content
Premium Trial:

Request an Annual Quote

German Team Classifies Lung Tumors With Methylation Profiles, Machine Learning Methods

NEW YORK – DNA methylation profiling on lung tumors — done in concert with machine learning methods — can help to discriminate between primary lung tumors and those that have landed there by metastases from head and neck cancers, new research suggests.

Although metastases to the lung can provide prognostic clues for individuals with head and neck squamous cell carcinoma (HNSCC), German researchers explained, lung tumors can also develop in parallel with such brain tumors, independent of metastasis. That can make it difficult to know whether lung squamous cell carcinoma (LUSC) tumors have arisen by metastasis, despite the importance of this distinction for an individual's care.

"Whereas distant metastases of HNSCC are mostly incurable and patients mainly receive only palliative chemotherapy or radiotherapy," the authors wrote online today in Science Translational Medicine, "patients with locally limited LUSC normally qualify for potentially curative therapy including lung lobectomy."

They reasoned that DNA methylation differences might provide clues for distinguishing the sources of these tumors, which often share histological, morphological, immunohistochemical, and even mutational features.

"The DNA methylation signatures of different tissue types are known to be quite specific, which has resulted in the recent development of promising algorithms that can characterize cancers of unknown primary tumors, brain tumors, and sarcomas according to their epigenetic signatures," the authors wrote. "On the basis of these results, we aimed to develop a DNA methylation-based machine learning classifier that facilitates the diagnostic differentiation of [head and neck squamous cell carcinoma] metastases to the lung from primary LUSC."

The researchers retrospectively profiled lung tumors from 408 individuals with a history of HNSCC, focusing in on 51 cases where it was possible to discern between HNSCC metastases to the lung or primary LUSC based on histopathological, molecular, and other available data.

With the help of epigenetic data for primary head and neck or lung cancer cases profiled for large efforts such as the Cancer Genome Atlas (TCGA) or Gene Expression Omnibus, the team focused in on thousands of cytosine methylation sites that varied between the cancer types, using those methylation differences to train three machine learning algorithms built on an artificial neural network framework.

After further tuning and tweaking those computational classifiers, the researchers demonstrated that their models could pick primary HNSCCs from primary LUSC tumors with more than 96 percent accuracy using publicly available methylation profiles for 279 tumors.

Returning to the 51 clinical samples on hand, meanwhile, the team found that both the neural network and support vector approaches showed increased positive predictive values compared to a random forest classifier, the team reported. Those results were further supported by independent analyses on samples from four LUSC patients profiled for the TCGA who had also been diagnosed with HNSCC.

"With this study, we successfully demonstrate that DNA methylation profiling in conjunction with machine learning solves the diagnostic problem of differentiating lung metastases of squamous cell carcinomas of the head and neck from primary lung cancers arising in patients with previous or simultaneous HNSCC," co-senior and co-corresponding authors Frederick Klauschen and David Capper, researchers based at the German Cancer Research Center and elsewhere, and their co-authors wrote.

Based on their findings so far, the authors suggested, "our approach may facilitate the reliable diagnostic differentiation of pulmonary metastases of HNSCC from primary LUSC to guide therapeutic decisions."