Dictionary learning for integrative, multimodal, and massively scalable single-cell analysis

This paper introduces 'bridge integration', a method to integrate single-cell datasets across different molecular modalities using a multiomic dataset as a molecular bridge. Mapping single-cell sequencing profiles to comprehensive reference datasets provides a powerful alternative to unsupervised analysis, but most reference datasets are built from single-cell RNA-sequencing data and cannot annotate datasets that do not measure gene expression. In bridge integration, each cell in the multiomic dataset constitutes an element in a 'dictionary' that is used to reconstruct unimodal datasets and transform them into a shared space, extending dictionary learning to the single-cell setting.

The procedure accurately integrates transcriptomic data with independent single-cell measurements of chromatin accessibility, histone modifications, DNA methylation, and protein levels. The authors further demonstrate that dictionary learning can be combined with sketching techniques to improve computational scalability, harmonizing 8.6 million human immune cell profiles from sequencing and mass cytometry experiments. Implemented in version 5 of the Seurat toolkit, the approach broadens the utility of single-cell reference datasets by extending reference mapping beyond scRNA-seq to single-cell epigenetic and proteomic data.