Bridge integration leverages tools from a subfield of representation learning known as ‘dictionary learning’, which are commonly used in image analysis33,34. The goal of dictionary learning is to find a representation of the input data as a weighted linear combination of individual basic elements. We show that dictionary learning has multiple potential applications for single-cell analysis. Our bridge integration procedure is enabled by treating each cell in a multi-omic dataset as elements of a dictionary that can be utilized to reconstruct single-modality datasets. Moreover, we demonstrate how the development of compact dictionaries via dataset sketching can dramatically improve the computational efficiency of large-scale single-cell analysis, and enable rapid integration of dozens of datasets spanning millions of cells.
