Skip to main content
Premium Trial:

Request an Annual Quote

Imaging Mass Cytometry Discovers Tumor Microenvironment Features Predicting Lung Cancer Outcomes

NEW YORK — Researchers from McGill University and elsewhere have developed an approach to predict from just a small tumor tissue sample which lung cancer patients will progress following surgery.

Specifically, they combined imaging mass cytometry with deep learning to analyze the tumor microenvironment of lung cancer samples from more than 400 lung adenocarcinoma patients. That environment, they noted, has been identified as a source of heterogeneity that affects treatment response and disease progression.

By characterizing the tumor microenvironment spatially and at the single-cell level, the researchers were able to uncover different cellular states and features associated with clinical characteristics such as survival. As they reported in Nature on Wednesday, they then used artificial intelligence to identify which of these features of the tumor microenvironment could predict disease progression with high accuracy.

"Overall, these data suggest that spatially resolved single-cell datasets may be highly valuable in the future to help to inform personalized peri-operative care plans to minimize toxicity for those destined to be cured, or to increase cure rates for those destined to recur," the researchers, led by co-senior authors Daniela Quail and Logan Walsh from McGill and Philippe Joubert of Laval University, wrote in their paper.

Using an imaging mass cytometry system from Fluidigm (now Standard BioTools), the researchers analyzed small tissue core samples from 426 patients with lung adenocarcinoma that were collected between February 1996 and July 2020. They used a 35-plex antibody panel to identify the various cellular components of their samples, including the cancer cells themselves as well as stromal cells and adaptive and innate immune cells. In total, the researchers detected more than 1.6 million cells and uncovered 14 distinct immune cell populations.

They focused in particular on the immune cell populations and whether they were associated with patients' clinical data. Mast cells, for instance, were linked to prolonged survival, though they were more common among nonsmokers and patients with early-stage disease. The researchers further noted ties between the frequencies of certain immune cells and particular clinical subgroups — CD4-positive helper T cells, for example, were enriched in samples from female patients, who tend to have better overall survival, while older patients had fewer intratumoral CD8-positive T cells.

At the same time, they explored how different cellular phenotypes in the tumor microenvironment were associated with survival, finding, for instance, that H1F1-alpha-positive neutrophils were linked to worse survival.

Looking at cellular neighborhoods — areas with similar local cell-type composition — the researchers further noted that various tissue architectures were linked to differences in survival. Cellular neighborhoods enriched in B cells, for instance, were significantly associated with survival, especially the CN-25 neighborhood, which was also enriched for CD4-positive helper T cells.

By applying a deep-learning approach, the researchers found that the spatial information they had generated could improve predictions of clinical outcomes. Their model, including spatial information, predicted progression with 95.9 percent accuracy, as compared to the 75 percent accuracy of the baseline score, they reported, and used a single 1 mm2 tumor core.

Further, the researchers analyzed a separate validation cohort of 60 patients with primary lung adenocarcinoma using imaging mass cytometry and found in this dataset that the model could predict progression with 94 percent accuracy.

The researchers traced the predictive abilities of their model to the combination of six markers: CD14, CD16, CD94, αSMA, CD117, and CD20. Together, these had an accuracy of 93.3 percent along with 95.6 percent precision and recall.

"Our findings represent an important advance over existing prediction tools that use clinical and pathological variables and may enable more effective utilization of a growing armamentarium of peri-adjuvant systemic therapies to improve cancer outcomes," the researchers wrote.