DeSide: A unified deep learning approach for cellular decomposition of bulk tumors based on limited scRNA-seq data

Cellular decomposition employing bulk RNA-sequencing (RNA-seq) has been consistently under investigation due to its high fidelity, ease of use, and cost-effectiveness compared to single cell RNA-sequencing (scRNA-seq). However, the intricate nature of the tumor microenvironment, and the significant heterogeneity among patients and cells have made it challenging to precisely evaluate the cellular composition of solid tumors using a unified model. In this work, we developed DeSide, a deep learning and single-cell decomposition method for solid tumors, to estimate proportions of cell types presented in tumor samples. Our new deep neural network (DNN) architecture considers only non-cancerous cells during the training process, indirectly calculating the proportion of cancerous cells. This approach avoids directly handling the often more variable heterogeneity of cancerous cells, and instead leverages scRNA-seq data from three different cancer types to empower the DNN model with a robust generalization capability across diverse cancers. Additionally, we used a new sampling method and filtering strategies to simulate the gene expression profiles (GEPs) of solid tumors, creating a carefully controlled training set that could be compared to the bulk RNA-seq data from The Cancer Genome Atlas (TCGA), a database of bulk RNA-seq data collected from cancer patients. Relying on limited yet diverse scRNA-seq data, our approach outperformed current methods in accurately predicting the celluar composition of samples from TCGA and an additional validation set. Furthermore, we demonstrated that the predicted cellular composition can be utilized to stratify cancer patients into different groups with varying overall survival rates. With increased availability of scRNA-seq data for various types of tumors, DeSide holds the potential for a more precise cellular decomposition model using bulk RNA-seq.