DeSide

The tumor microenvironment is a complex mixture, with diverse cell type interactions. Due to the variability of expression patterns within and across cell types, accurately inferring cell type abundance in the mixture can be challenging. Machine learning faces a dilemma: train specific models for each tumor type with limited data or one model for all tumors with vast data. We present methods to refine data simulation with real tumor data guidance to generate a modest training set and tailor a unified deep learning model to bypass the issue of highly variable gene expression in cancer cells. Providing precise cell abundance estimation, our approach sheds light on cell type interactions in the tumor microenvironment and potentially improves clinical diagnosis.