Abstract
Current single-cell profiling technologies enable the capture of multiple cellular modalities, providing valuable insights into complex biological systems. While a substantial amount of single-cell multimodal data has been generated and accumulated, most of these datasets are unpaired, characterized by distinct feature spaces and a lack of cell-wise correspondence. The absence of explicit linkages between modalities poses a fundamental challenge for data integration and interpretation. To address this, we introduce SuperMap, a statistical learning method designed for the integrative analyses of unpaired multimodal data. SuperMap directly learns cross-modal mappings from unpaired data to effectively bridge and link different modalities, facilitating a variety of downstream analysis tasks. Comprehensive benchmarking and real-world applications demonstrate the superior performance of SuperMap in enhancing cell-type identification, improving diagonal integration, enabling regulatory analysis, and revealing epigenomic priming events to specify cell differentiation directions for trajectory inference.