Harnessing multi-modal and multi-omics data integration to decipher therapeutic response in non-small cell lung cancer

Stanford University, Stanford CA

Investigators

Abstract

Project Abstract Lung cancer is the leading cause of cancer-related deaths worldwide, with non-small cell lung cancer (NSCLC) accounting for 80% of all lung cancer cases. Treatment options include surgery, radiation, chemotherapy, targeted therapy, and immunotherapy. However, these treatments provide clinical benefits to only a subset of patients due to treatment resistance stemming from clonal diversity and cell-state plasticity. Recent studies have revealed that the tumor microenvironment, comprised of immune cells and non-immune stromal elements, can enhance tumor cell stemness, proliferation, and metastasis through intricate intercellular communications. To characterize the spatial interactions between tumor cells and their microenvironment, technical advancements have been made in profiling the spatial cellular architectures and signaling interactions with high-dimensional, multi-omics, and multi-modal biomedical data. Nevertheless, effective computational approaches for integrating information from these datasets are still lacking. This presents three challenges: (1) relying on a single data modality captures only partial disease characteristics, thus limiting precise patient stratification and outcome prediction; (2) technologies using high-throughput sequencing or multiplexed imaging are still expensive, requiring specialized equipment, and have not yet been integrated into routine diagnostics; (3) large-scale tumor profiling is constrained by platform throughput and tissue availability. As a result, how the cell-type diversity and their spatial architecture is associated with clinical outcomes has not been completely understood. I hypothesize that harnessing computational approaches to integrate multi-modal and multi-omics biomedical datasets can improve personalized diagnosis and treatment. To demonstrate the benefits of combining multi-modal data, I recently developed GBM360, a machine learning framework that utilizes histology images to infer the spatial distribution of malignant cells and prognosis in glioblastoma. By integrating single-cell RNA-seq, spatial transcriptomics, and histology images, I phenotypically analyzed 40 million tissues spots from 410 patients. The results link spatial cellular architectures to patient prognosis. However, the study solely focused on the analysis of tumor cells, ignoring the impact of immune cells and non-immune stromal elements to clinical outcomes. Additionally, whether the approach can be extended to other tumor types remains unknown. The proposed study aims to address three key issues: (1) identifying clinically relevant spatial cellular architectures contributing to therapeutic response; (2) developing an image-based digital cytometer to computationally reconstruct the spatial cellular landscape from histology images; (3) creating a unified machine-learning framework for integrating tumor microenvironmental information from multi-modal data to improve prognostic predictions and treatment allocation.