Comprehensive Enumeration of Cancer Stem-like Cell Heterogeneity Using Deep Neural Network

Abstract

Cancer stem cells (CSCs), a distinct subpopulation within tumors, are pivotal in driving treatment resistance and tumor recurrence, posing substantial challenges to conventional therapeutic strategies. Precise quantification and profiling of these cells are essential for improving cancer treatment outcomes. We present ACSCeND, an advanced deep neural network model accompanied by a robust workflow, specifically developed to quantify cellular compositions from bulk RNA-seq data, enabling accurate CSC profiling. By integrating bulk RNA-seq data with insights derived from single-cell RNA-seq datasets, ACSCeND effectively captures the diversity and hierarchical organization of tumor-resident cell states, alongside cell-specific gene expression profiles (GEPs). Compared to current tissue deconvolution models, ACSCeND exhibits superior performance, achieving significantly higher Concordance Correlation Coefficient (CCC) values and lower Root Mean Square Error (RMSE) across various pseudobulk and real-world bulk tissue samples. Application of ACSCeND to TCGA and PRECOG datasets reveals a strong association between CSC abundance and poorer disease-free survival outcomes, underscoring the clinical relevance of CSCs in cancer progression. Furthermore, cell-specific GEPs for distinct CSC states unveil novel molecular signatures and illuminate the origins of CSC-driven tumor heterogeneity. In summary, ACSCeND provides a powerful, scalable platform for high-throughput quantification of cellular compositions and distinct potency states within normal tissues as well as highly heterogeneous tissues, such as tumors.