Jointly aligning cells and genomic features of single-cell multi-omics data with co-optimal transport

Abstract

The availability of various single-cell sequencing technologies allows one to jointly study multiple genomic features and understand how they interact to regulate cells. Although there are experimental challenges to simultaneously profile multiple features on the same single cell, recent computational methods can align the cells from unpaired multi-omic datasets. However, studying regulation also requires us to map the genomic features across different measurements. Unfortunately, most single-cell multi-omic alignment tools cannot perform these alignments or need prior knowledge. We introduce O_SCPLOWSCOOTRC_SCPLOW, a co-optimal transport-based method, which jointly aligns both cells and genomic features of unpaired single-cell multi-omic datasets. We apply O_SCPLOWSCOOTRC_SCPLOW to various single-cell multi-omic datasets with different types of measurements. Our results show that O_SCPLOWSCOOTRC_SCPLOW provides quality alignments for unsupervised cell-level and feature-level integration of datasets with sparse feature correspondences (e.g., one-to-one mappings). For datasets with dense feature correspondences (e.g., many-to-many mappings), our joint framework allows us to provide supervision on one level (e.g., cell types), thus improving alignment performance on the other (e.g., genomic features) or vice-versa. The unique joint alignment framework makes O_SCPLOWSCOOTRC_SCPLOW a helpful hypothesis-generation tool for the integrative study of unpaired single-cell multi-omic datasets. Available at: https://github.com/rsinghlab/SCOOTR.