Optimal transport reveals unresolved cells in single cell multi-omics

Abstract

Single-cell multi-omics methods align cells across modalities, integrate them into one embedding and transfer labels between them. All three steps depend on a match between cells in two modalities, but this match usually stays hidden inside each tool. Its errors are hidden too, and every cell looks equally certain. Here we make the match explicit. For each pair of datasets, we build a soft coupling matrix with optimal transport and use it in three ways beyond estimating the match itself: to score how uncertain each cell's match is, to upgrade existing integration and label-transfer tools without retraining them, and to decide which cells to set aside. Across 78 dataset pairs spanning six modality combinations, we evaluated ten integration and ten label-transfer tools using matched random removal as the control. We developed an ensemble coupling achieving the most accurate recovery of true cell pairs. Feeding couplings back into the tools improved label transfer clearly but improved integration only in how well the modalities mix. The ambiguity score, which we computed as a group-vote margin, gave the larger and more consistent gain in both tasks. Ambiguous-flagged cells matched worse in almost every dataset and removing them beat random removal in 98% of integration tests and 91% of label-transfer comparisons. These hard-to-integrate cells are mostly among rare, transitional and poorly measured populations. Making correspondence explicit therefore provides a practical way to distinguish reliable cross-modality results from uncertain ones while flagging cells that might have important roles in development and disease.