Skip to main content
Premium Trial:

Request an Annual Quote

Advanced Cancer Outcomes Predicted Using Model Based on Circulating Tumor DNA Metrics

NEW YORK – A team from Genentech, Foundation Medicine, and elsewhere has come up with a machine learning model for bringing together a range of circulating tumor DNA (ctDNA) metrics to predict treatment response and related survival patterns in advanced lung cancer patients — an approach that is expected to inform future clinical trials looking at candidate immunotherapy drugs or treatment combinations.

The study appeared in Nature Medicine on Thursday.

Starting with ctDNA samples collected over time for 466 stage IV non-small cell lung cancer (NSCLC) patients participating in a randomized Phase III study of several immune checkpoint inhibitor-chemotherapy combinations known as Impower150, the researchers used machine learning to bring together ctDNA features and clinical data for the patients, focusing on multiple ctDNA metrics linked to overall survival (OS).

"There are diverse approaches used in the literature to summarize ctDNA levels and integrate ctDNA features for association with clinical outcomes, as well as the open question regarding which on-treatment time points may be optimal for longitudinal ctDNA analyses," co-senior and co-corresponding authors Katja Schulze, director of oncology biomarker development and medical affairs at Genentech, and David Shames, Genentech's executive director and senior fellow in oncology biomarker development, explained in an email.

"To our knowledge, this is the first study to systematically evaluate the use of longitudinal ctDNA dynamics across a large, randomized Phase III clinical trial," Schulze and Shames added.

At the moment, imaging methods such as magnetic resonance imaging or computerized tomography are typically used as surrogate markers to gauge response to treatment, they explained, although there is mounting evidence that the survival benefits associated with immunotherapy-based treatments do not track as closely with imaging-based treatment responses as they do in patients receiving cytotoxic treatments such as chemotherapy.

Consequently, investigators are on the hunt for additional early-disease markers that can provide a window into longer-term treatment response and survival, particularly when it comes to clinical trials evaluating new immunotherapy candidates or treatment combinations.

"Because of this lack of correlation between surrogate measures of drug efficacy and OS, oncology drug trials often depend on OS as a primary endpoint. This means that trials can take many years to complete," the study's authors wrote, adding that "there is an important need to evaluate immunotherapy drug efficacy early in the course of therapy using alternative methods that are better associated with OS."

With that in mind, the researchers used FoundationOne Liquid CDx and custom panel assays to assess mutations across hundreds of genes in more than 1,900 ctDNA samples collected at five time points from IMpower150 trial participants with metastatic, nonsquamous NSCLC.

By plugging these dynamic ctDNA metrics into a machine learning model, the team was able to find prognostic clues in ctDNA samples collected relatively early in the treatment process. In particular, ctDNA information found in blood samples collected on the first day of patients' third treatment cycle — around six weeks into treatment — could be used to predict OS or disease recurrence risk, regardless of imaging-based treatment response predictions.

Along with validation analyses done using data for NSCLC patients from another Phase III clinical trial — which supported the notion that the ctDNA-based model can identify high-risk patients who might benefit from alterative treatment — findings from the investigators' simulation studies suggested that the ctDNA-based model could perform better than imaging-based approaches when evaluating predicted survival outcomes associated with new treatments in earlier stage Phase I or Phase II trials.

"Our findings suggest that on-treatment ctDNA measurements can complement radiographic imaging to inform on whether patients are responding to therapy, which has implications for patient management during the course of their cancer care," Schulze and Shames said. "Our findings also suggest that ctDNA might be useful for identifying novel molecules that might provide survival benefit to patients in early Phase I/II drug development settings."

The pair noted that ctDNA metrics that appeared particularly informative included the number of variants found in the gene panels assayed, the average number of tumor molecules in a given amount of plasma, and the total amount of cell-free DNA extracted from a blood plasma sample.

Even so, the authors noted that more research will be needed to understand how ctDNA assay selection might impact the precise ctDNA metrics considered and the resulting prediction model.

"Our work suggests exciting next steps, for example, incorporating other circulating biomarkers into our model and assessing whether they are superior to ctDNA for predicting outcome," Schulze and Shames explained, noting that "prospective validation of our findings is necessary to bring such approaches into routine clinical practice."