Single-cell assay for transposase-accessible chromatin using sequencing (scATAC-seq) can determine cell types, states and differentiation trajectories within the heterogeneous tissues. However, it remains challenging to accurately distinguish tumor subpopulations from the scATAC-seq assay. Here, we present a novel multimodal matrix factorization method called MAAS, which integrates chromatin accessibility, copy number variations and single-nucleotide variants solely from scATAC-seq data to identify functional tumor subpopulations with genetic variability. Systematic benchmarking of MAAS demonstrated its superior accuracy (>0.9) and robustness against changed number of cells and subpopulations, compared to state-of-the-art tools for identifying cell subpopulations. When applied to a glioma scATAC-seq dataset, MAAS revealed previously obscured subsets of cells associated with worse survival and higher risk of hypermutation, hidden by copy number variations. In B-cell lymphoma and renal cancer, MAAS successfully deconvoluted progressive tumor subpopulations linked to poorer prognosis and distinct drug responses. In summary, MAAS identifies biologically and clinically pertinent tumor subpopulations by directly integrating genetic and epigenetic features from scATAC-seq data, thus expediting the discovery of potential therapeutic targets.