Photo Credit: iStock.com/Elena Merkulova

A new study employing novel single-cell multimodal data revealed key insights into the tumor microenvironment predictive of NSCLC treatment response.

Leveraging advanced deep learning techniques and multimodal data has uncovered novel insights into the tumor microenvironment (TME) of non-small cell lung cancer (NSCLC), potentially enhancing patient stratification and treatment selection, according to findings published in Science Advances.

NSCLC is the most common form of lung cancer, responsible for over 80% of cases, and a leading cause of cancer-related deaths worldwide. “Despite the advent of immune checkpoint inhibitors, their efficacy is limited to 27 to 45% of patients,” wrote corresponding author Olivier Gevaert, PhD, MS, Stanford Center for Biomedical Informatics Research, and colleagues. “Identifying likely treatment responders is essential for optimizing healthcare and improving quality of life.”

Multiphase Process Yields Numerous Results

Based on the hypothesis that integrating tumor microenvironmental information from multimodal biomedical datasets can guide the selection of effective treatment regimens, the team began their research by generating and evaluating multiplex immunofluorescence (mIF) imaging, histopathology, and RNA sequencing data from NSCLC tissues. Analysis led to the spatial organization characterization of 1.5 million cells based on expression levels for 33 biomarkers and revealed distinct immune cell compositions in different histological types of NSCLC. For example, squamous cell carcinoma exhibited higher proportions of B cells and monocytes, while adenocarcinoma showed more T helper cells and dendritic cells, according to findings.

The team then identified specific spatial cellular neighborhoods that influence tumor progression. The authors found that tumors with higher proportions of lymphocyte-enriched neighborhoods showed extended progression-free survival, while macrophage-enriched neighborhoods were associated with poorer outcomes. In terms of immunotherapy response, tumors from responders had a significantly higher proportion of cytotoxic T (Tc) cells, which were spatially closer to dendritic cells, indicating a robust immune response.

Next, the team developed and applied NucSegAI—a deep learning model designed for automated nuclear segmentation and cellular classification in histology images—to 119 whole-slide histology images from 115 patients with NSCLC treated at Stanford Medical Center.

“In total, we detected 19.7 million tumor cells, 12.3 million lymphocytes, 3.3 million macrophages, 0.7 million vascular cells, and 9.6 million fibroblasts,” the authors reported.

By comparing the histology-based phenotype proportions of lymphocytes with the immune cell-type proportions estimated from deconvolution of sample-matched tissue-level RNA-seq data, the team discovered that a histology-based cytotoxic T lymphocyte (CTL) score, derived from lymphocyte phenotypes, was a strong predictor of immunotherapy response, outperforming traditional PD-L1 immunofluorescence.

Lastly, the researchers investigated whether integrating results from the three data modalities could improve predictive performance.

“To this end, we selected immunotherapy-treated patients (n=20) with all three data modalities available, and the CTL scores independently obtained from each modality were summed to generate a ‘fusion score’ for each patient,” the authors explained. Receiver operating characteristic analysis showed:

• mIF images alone achieved an area under the curve (AUC) score of 0.89 in identifying treatment responders;
• Histology images alone achieved an AUC of 0.82; and
• RNA-seq data alone achieved an AUC of 0.90.

“In comparison, the fusion score achieved the highest AUC of 0.94, outperforming any single modality,” the researchers stated. “These results indicate that combining microenvironmental modeling data from mIF imaging, RNA-seq, and histology may improve patient stratification.”

Moving Towards Personalized Cancer Treatment

“Our methods offer an automated and scalable approach to characterizing clinically relevant spatial cellular phenotypes from histology images,” the authors wrote.

The integration of mIF, RNA-seq, and histology data offers a comprehensive approach to understanding the complexities of the TME in NSCLC tissues, according to the authors.

“Developing effective computational methods to integrate microenvironmental data from multimodal, multiomics biomedical datasets can guide the selection of personalized therapeutic approaches for individual patients, with the potential to improve precision medicine and cancer treatment strategies,” the team concluded. “Future work will evaluate the model’s generalizability across additional cancer types beyond NSCLC.”