Machine Learning Yields Lung Cancer Test

Patients with lung cancer may have better odds of survival if the disease is caught early, but developing screening tools that are both robust and widely adopted has been an obstacle. That's the challenge that motivated the development of a test dubbed “lung cancer likelihood in plasma” (Lung-CLiP), a liquid biopsy–based technique designed using machine learning.

With Lung-CLiP, lung cancer can be detected with sensitivity and specificity that can be adjusted according to the goal. For example, higher specificity would be needed if Lung-CLiP were used as a stand-alone test to minimize false positives, but if patients with positive Lung-CLiP tests could be referred for follow-up low-dose CT screening, some specificity could be sacrificed for higher sensitivity. The sensitivity for stage I disease was 63% at 80% specificity, values that may be appropriate in the latter scenario.

“We believe that developing a blood-based test could increase the number of at-risk patients screened,” says Maximilian Diehn, MD, PhD, of Stanford University in California, one of the authors of the article describing the test (Nature 2020;580:245–51).

Because of Lung-CLiP's ability to identify early-stage disease and potential to increase screening rates, Christian Rolfo, MD, PhD, MBA, of the University of Maryland School of Medicine in Baltimore, says, “This is a very, very important article,” noting the study's methodologic rigor.

As a proof of concept that a liquid-biopsy test might be effective for early detection of lung cancer, the study's authors first ascertained that patients with stage I disease often have circulating tumor DNA (ctDNA). Then, they sought to distinguish tumor-derived mutant DNA in the plasma from DNA with mutations arising from clonal hematopoiesis, a natural consequence of aging. To do this, they sequenced both leukocytes and plasma from the same patients and relied on observations about characteristics of the mutations found in the patients' ctDNA. For example, lung cancer–derived mutant DNA fragments were shorter on average than those from blood cells.

With this information, the authors applied machine learning to a training cohort of patients with early-stage lung cancer and high-risk controls in whom low-dose CT scanning didn't detect the disease. The variables identified were used to develop an algorithm that computes the probability that a blood sample contains DNA from lung cancer. To validate the approach, the investigators prospectively evaluated Lung-CLiP in an independent patient cohort from a different institution, a step Rolfo cites as a major strength of the study.

“We were very pleased to see that the performance in the validation cohort was statistically identical to the performance in the training cohort, suggesting that the performance of Lung-CLiP is robust and reproducible,” Diehn says.

Follow-up studies will dig into discoveries made during Lung-CLiP's development: For example, in patients with early-stage disease, the correlation between ctDNA levels and survival indicates that the test may have prognostic value. To definitively determine the test's clinical utility, Diehn says a randomized controlled trial will ultimately be required. “While our results are very encouraging, being able to detect cancer at an early stage doesn't necessarily prove that use of a test will save lives,” he says.

Assessing the clinical use of Lung-CLiP in a larger study is essential, agrees Hilary Robbins, PhD, MHS, MSPH, of the International Agency for Research on Cancer in Lyon, France. “They've done commendable work here,” she says. “But it's a very long and challenging road to translate these assays into clinical practice.” –Nicole Haloupek

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