1196 Unveiling immune evolution in lung pre-cancer through graph contrastive learning of spatial transcriptomics

<h3>Background</h3> Exploring the tumor immune microenvironment from a pre-cancer perspective using spatial transcriptomics (ST) provides critical insights into the early changes and interactions that precede tumor formation. This understanding is vital for developing immunoprevention strategies that modulate the immune landscape to intercept cancer progression. We developed CoCo-ST (Comparing and Contrasting Spatial Transcriptomics), a method that uses a background dataset of normal tissue to identify distinctive patterns in precancerous tissues. By identifying these early molecular and cellular changes, CoCo-ST informs targeted immunotherapy interventions, potentially enhancing immune surveillance and improving early intervention outcomes. <h3>Methods</h3> Fresh normal lung and lung tumor tissue samples collected from mice were processed using both single-cell sequencing and spatial transcriptomics technologies from 10x Genomics. Using our CoCo-ST method, we constructed two weighted graphs for the spatial transcriptomics (ST) datasets—one for the background (normal tissue) and one for the target (tumor tissue). These graphs captured the local structures within each dataset. We then derived contrastive feature representations by comparing and contrasting the local variances between the background and target graphs. This was accomplished by assessing the differences between their respective local total scatter matrices, enabling us to identify distinctive patterns and changes specific to the tumor microenvironment. <h3>Results</h3> We demonstrated the effectiveness of CoCo-ST’s contrastive feature representations across multiple downstream analysis tasks, including spatial transcriptomics (ST) data visualization, spatial domain identification, and spatial trajectory inference. CoCo-ST effectively identifies feature representations that robustly discriminate various spatial structures within the target tissue, closely aligning with pathologist-annotated regions. In contrast, spatial clustering based on components from existing methods, such as Seurat (PCA), STUtility (NMF), NSF, and MEFISTO, failed to effectively identify the hyperplasia hotspot region annotated by the pathologist. For trajectory inference, CoCo-ST’s contrastive components successfully distinguished the different spatial domains in the tissues and identified a trajectory from normal lung tissue to adenoma and adenocarcinoma, consistent with known mouse tumorigenesis biology. By elucidating these early molecular and spatial changes, CoCo-ST can inform immunoprevention strategies to prevent cancer progression and enhance immune surveillance, ultimately leading to more precise and effective treatments, thereby improving patient outcomes. <h3>Conclusions</h3> We introduced an ST feature representation method that offers significant advantages, particularly in scenarios involving the analysis of multiple ST data sets. It effectively identifies biologically-relevant, unique spatial structures in a target ST data set while mitigating the influence of dominant high-variance spatial structures that are common to both target and background ST data sets. <h3>Ethics Approval</h3> All animal experiments were conducted following MD Anderson Institutional Animal Care and Use Committee–approved protocols (approval number 00001217-RN03).