SCALEMAP swaps cross-modal anchor search for self-mapping, integrating 1.5 million single-cell profiles in under five minutes
Synopsis
In single-cell multi-omics, modalities such as RNA and protein are often measured in different cells, so there is no cell-level correspondence and alignment relies on predefined feature links to find cross-modal anchor cells, which becomes costly at atlas scale. SCALEMAP changes the alignment signal: each cell is translated into the other modality and compared with the linked features observed in that same cell, giving every cell its own positive target without first searching for a partner cell. On RNA–protein and RNA–ATAC benchmarks its integration quality matches or exceeds leading methods at substantially lower runtime and memory; on the COMBAT atlas it integrated more than 830,000 cells in 3.3 minutes.
Interpretation
Cross-modal alignment no longer requires searching observed cells for paired anchors; each cell uses its own translation as the alignment target. Earlier weak-linkage methods first inferred cross-modal anchor cells through predefined feature links and then used those anchors to guide alignment, and anchor search and refinement grew computationally burdensome as cell counts rose. On the COMBAT atlas CITE-seq data, across six dataset sizes from 2% to 100%, SCALEMAP had the shortest runtime at every scale and integrated the full dataset in 3.3 minutes; BindSC exceeded the memory limit at moderate sizes, MaxFuse and GLUE ultimately exceeded the 12-hour limit, and scMODAL needed 87.5 minutes for the full dataset under default settings.
The alignment signal comes from each cell's own cross-modal translation rather than from a cell in the other modality judged similar. Each input cell is decoded into the other modality, the corresponding linked features are extracted through the predefined feature-link table, and they form a positive pair with the linked features observed in the same cell. Training runs in a multi-input, multi-output autoencoder: encoders map both modalities into a shared latent space, and modality-specific decoders both reconstruct their own modality and generate a translated profile in the other; a contrastive loss requires the same-cell positive pair to outscore translated vectors from other cells in the minibatch, optimized jointly with reconstruction, latent consistency and adversarial modality mixing.
Integration quality matches or exceeds leading methods while computational cost is markedly lower. Across six RNA–protein datasets, SCALEMAP achieved the highest overall score, followed by scMODAL, MaxFuse and BindSC; BindSC scored on average 11.3% and 8.0% lower on biological conservation and batch correction. On the full COMBAT dataset, relative to scMODAL, SCALEMAP scored 25.5% higher on ARI, 34.6% higher on batch LISI and 23.1% higher on 1-FOSCTTM; after reducing scMODAL training from 10,000 to 500 steps to match runtimes, SCALEMAP scored 76.4% higher on ARI, 199.6% higher on batch LISI and 60.9% higher on 1-FOSCTTM.
Perspective
Evaluation covers six RNA–protein datasets (Bone Marrow CITE-seq, PBMC TEA-seq, a COMBAT subset, PBMC ASAPseqCtrl, PBMC ASAPseqStim and PBMC CITE-seq) and three RNA–ATAC datasets, with scale experiments run on a 32-core Intel CPU under a 12-hour runtime limit and a 100 GiB memory limit.
Self-mapping still relies on predefined feature links to constrain cross-modal mapping and currently uses a fixed set of linked features; the authors propose extending it to learn link-specific weights or refine the feature comparison map under biological constraints to accommodate variation in link informativeness across cell states.
