We next performed quantitative benchmarking of both multi-omic integration methods (bridge integration, Cobolt, multiVI), and also evaluated ‘bridge-free’ methods (Canonical Correlation-based Integration, LIGER), which perform integration on the basis of gene activity scores. For this analysis, we began by splitting the 10x multi-omic bridge dataset into two groups. In one group, we treated the scRNA-seq and scATAC-seq data as if they were from separate experiments, representing a benchmark dataset for integration where ground-truth correspondences were known. The second group of cells was used as a multi-omic bridge dataset. After aligning cells across modalities, we calculated the Jaccard similarity metric between each scATAC-seq cell and its matched scRNA-seq counterpart, representing a quantitative metric that does not rely on discretized cell labels.
