Commentary
Video
Author(s):
Katharina Hoebel, MD, PhD discusses the efficacy of incorporating spatial interactions within predictive models, in the realm of oncology.
Katharina Hoebel, MD, PhD, research fellow, Harvard Medical School, Dana-Farber Cancer Institute, discusses research involving spatial interactions to develop a predictive model to correlate expression patterns with the prognosis of non–small cell lung cancer (NSCLC).
Hoebel and colleagues hypothesized that modeling these spatial interactions in the could enhance survival predictions in NSCLC, and findings from their research presented at the 2024 AACR Annual Meeting validated that theory, she says. The research utilized biopsy samples from patients with NSCLC, and they were stained using a targeted multiplex immunofluorescence assay. Data from the samples were used to develop a deep-learning model that leverages cellular neighborhood graphs, which can identify expression patters that correlate with NSCLC prognosis. Hoebel explains that this model outperformed other baseline measures—including PD-L1 tumor proportion score (TPS) and immune marker density–based models—that do not account for spatial interactions.
Employing interpretability methods, 3 distinct clusters enriched in PD-L1 stood out, she continues. One cluster was enriched in PD-L1–positive tumor cells, and 2 clusters exhibited enrichment in PD-L1–positive immune cells, she says.
When the latter 2 clusters were enriched with PD-L1–positive immune cells, this correlated with favorable survival outcomes, she continues.In 1 cluster, the PD-L1–positive immune cells were dominant, and in the other, the PD-L1–positive immune cells were grouped with other immune cells, including CD8- and PD-1–positive cells. When PD-1–positive immune cells were mixed in this cluster, this correlated with superior survival, Hoebel concludes.