All Times EDT
Keywords: precision medicine, precision oncology, drug sensitivity prediction, cancer proteomics, Bayesian models, graphical regression
A systemic approach in precision medicine has been to bridge anticancer pharmacological data to large-scale multi-layered molecular tumor profiles using cell lines as proxies for cancer patients. However, samples from different tumor micro-environments in the two model systems may exhibit distinct patterns of molecular activities in general. Specifically, the architecture of cancer modulation through cumulative effects from multiple interacting genes in functional/signaling pathways may vary across model systems. In this paper, we attempt to address these challenges by developing a multi-level Bayesian analytical framework called TransPRECISE: Translational Personalized Cancer-specific Integrated Network Estimation. TransPRECISE uses Bayesian graphical regression models to infer on cancer-specific pathway circuitry by establishing weighted connections between the genes in the pathway based on the posterior inclusion probabilities of the protein-protein interactions. These cancer-specific networks can then be de-convolved to a sample-specific level by identifying node (gene / protein) labels as neutral, activated or suppressed, which can further be summarized to neutrality, activation or suppression scores and statuses of each pathway for each sample. We use TransPRECISE to analyze proteomic data for 640 cell lines from 16 lineages and 7714 patients from 31 tumor types. Through pan-cancer analysis, we investigate differential and conserved aspects of cancer-specific pathway networks across model systems and cancer lineages, identify matching avatar cell lines personalized to the patient tumor profiles and train Bayesian additive regression tree (BART) models based on cell lines for predicting drug sensitivity in patients. Finally, we also developed an R Shiny app for aiding seamless visualization of results and possible future investigations into our findings.