Scalable Multi-omics Diagonal Integration from Self-Mapping

Diagonal integration seeks to align single-cell multi-omics datasets in which distinct modalities, such as RNA and protein, are measured in different cells. Unlike horizontal integration, these datasets lack direct cell--cell correspondences and complete feature mappings across modalities. Leading methods for weak-linkage integration often infer anchor cells between modalities by searching for cross-modal similarities through predefined feature links, then use these anchors to guide alignment. Their required matching or anchor refinement can become computationally burdensome at atlas scale. Here, we introduce a novel self-mapping technique, a mechanism that aligns modalities by mapping each computationally decoded cross-modality omics profile prediction to its own observed inputs. We implement this strategy in SCALEMAP, a scalable, modality-agnostic method for diagonal integration. Across transcriptomic, epigenomic and proteomic benchmarks, SCALEMAP matches or outperforms leading alternatives while requiring substantially less runtime and memory. In large-scale experiments, SCALEMAP integrates more than 1.5 million multi-omic profiles in less than 5 minutes with under 20 GB peak memory, which enables rapid synthesis of massive single-cell multi-omics resources into unified biological representations.