NLM DIR Seminar Schedule

Computational methods to study interactions between tumors and their micro and macro environments

Tumor progression is shaped not only by cancer-cell-intrinsic changes but also by environmental and lifestyle factors. They actively remodel both the local tissue surrounding the tumor and the body's systemic environment to drive aggressive cancer growth and metastasis. Therefore, characterizing the regulatory roles of environments is critical to understand the mechanism of tumor progression. To this end, we developed three consequent approaches that characterize these relationships from multiple patients’ bulk data sets integrated with spatial transcriptomics. Each method addresses a distinct but complementary aspect of tumor progression.
First, GeneSigNet, is designed to identify causality links between biological processes and mutagenic processes. It constructs a weighted-directed network from gene expression and mutational signature exposures, using sparse partial correlation and a partial higher-moment strategy. In breast and lung adenocarcinoma, it recovered known relations, such as the effect of homologous recombination deficiency on clustered APOBEC mutations and the immune response to smoking.
Second, TranNet infers a sparse weighted bipartite graph from the gene expression of tumor-adjacent control samples to that of matched tumors. This identifies predictors, which are potential regulators of the normal-to-tumor transition. Applied to five TCGA cancer types, the predictors were enriched in G-protein coupled receptor signaling, cell-to-cell communication, immune processes, and cell adhesion. Their targets were enriched in tissue remodeling (including EMT), immune response, and proliferation. These findings suggest that adjacent normal tissue carries information on environmental, stromal, and microenvironmental influences on tumor progression.
Third, SpacePhenotyper transfers the predictive relationship between bulk gene expression and clinical phenotypes to spatially resolved transcriptomics. It builds an “Eigen-Gene” from principal directions of bulk data, transforms it into an “Eigen-Patient” marker vector, and scores spatial spots by similarity measure. It outperformed existing methods on simulated data. On real breast and colorectal tumors, its predictions matched pathologist annotations, and hazard was higher in invasive than in in-situ regions. It also revealed that predicted drug response (residual cancer burden) differs between in-situ and invasive regions, and within DCIS, highlighting the importance of early intervention.
Together, these methods provide a multi-scale framework for linking environmental factors to tumor-related gene regulation and clinical phenotypes. They offer new insights into the tumor and environment relationship, mutagenic processes, and spatial heterogeneity, with potential applications in early detection, prognosis, and treatment.

TBD

TBD

TBD

TBD