Abstract
Recent experimental advances have enabled high-throughput single-cell measurement of gene expression, chromatin accessibility and DNA methylation. We previously used integrative non-negative matrix factorization (iNMF) to jointly learn interpretable low-dimensional representations from multiple single-cell datasets using dataset-specific and shared metagene factors. These factors provide a principled, quantitative definition of cellular identity and how it varies across biological contexts. However, datasets exceeding 1 million cells are now widely available, creating computational barriers to scientific discovery. For instance, it is no longer feasible to use the entire available datasets as inputs to implement standard pipelines on a personal computer with limited memory capacity. Moreover, there is a need for an algorithm capable of iteratively refining the definition of cellular identity as efforts to create a comprehensive human cell atlas continually sequence new cells.
To address these challenges, we developed an online learning algorithm for integrating massive and continually arriving single-cell datasets. We extended previous online learning approaches for NMF to minimize the expected cost of a surrogate function for the iNMF objective. We also derived a novel hierarchical alternating least squares algorithm for iNMF and incorporated it into an efficient online algorithm. Our online approach accesses the training data as mini-batches, decoupling memory usage from dataset size and allowing on-the-fly incorporation of new data as it is generated. The online implementation of iNMF converges much more quickly using a fraction of the memory required for the batch implementation, without sacrificing solution quality. Our new approach enables factorization of 939489 single cells from 9 regions of the mouse brain on a standard laptop in ∼ 30 minutes. Furthermore, we construct a multi-modal cell atlas of the mouse motor cortex by iteratively incorporating seven single-cell datasets from three different modalities generated by the BRAIN Initiative Cell Census Network over a period of two years.
Our approach obviates the need to recompute results each time additional cells are sequenced, dramatically increases convergence speed, and allows processing of datasets too large to fit in memory. Most importantly, it facilitates continual refinement of cell identity as new single-cell datasets from different biological contexts and data modalities are generated.