Data-Driven Understanding of the Single-Cell Space in Tumor Microenvironment

Recent advances in single-cell sequencing have revealed heterogeneous cell types in the non-cancerous cells infiltrating tumors, which affects cancer progression and therapy responses. The integration of multiple scRNA-seq datasets across tumors can reveal common cell types and states in the tumor microenvironment (TME). We developed a data-driven framework, MetaTiME, which uses machine learning on millions of TME single cells, to learn "meta-components" that can interpret independent gene expression variations representing cell types, cell states, and signaling activities. The meta-components depict functionally distinct gene programs and reveal TME cell states and their gene regulation with higher granularity. Following the idea of transfer learning, by projecting onto the MetaTiME space, we create a tool that annotates cell states and signature continuums for TME scRNA-seq data. Overall, MetaTiME learns data-driven meta-components that depict cellular states and gene regulators useful for studies in tumor immunity and cancer immunotherapy.