Scalable joint non-negative matrix factorization for paired single cell gene expression and chromatin accessibility data.

Single-cell multi-modal technologies provide powerful means to simultaneously profile cellular states. These are now being employed to study gene regulatory mechanisms in a variety of biological systems. Tailored computational methods for integration and analysis of these data are much needed, with desirable properties in terms of efficiency-to cope with high dimensionality of the data, interpretability-for downstream biological discovery and hypothesis generation, and flexibility-to easily incorporate future modalities. Existing methods cover some but not all of the desirable properties for effective integration and analysis of these data. Here, we present a highly efficient method, q-intNMF, for representation and integration of single-cell multi-modal data using joint non-negative matrix factorization, which can facilitate discovery of linked regulatory topics in each modality. We provide thorough benchmarking using large publicly available datasets against five popular existing methods. q-intNMF performs comparably against the current state-of-the-art methods across a range of metrics. Additionally, q-intNMF provides advantages in terms of computational efficiency and interpretability of discovered regulatory topics in the original feature space. We illustrate this enhanced interpretability in providing insights into cell state changes associated with Alzheimer's disease. q-intNMF is available as a Python package with extensive documentation and use cases at https://github.com/wmorgans/quick_intNMF.