Single-cell multi-omics integration changes its starting point: factorizing cell similarities instead of rebuilding molecular measurements

Mosaic single-cell multi-omics integration can now avoid reconstructing molecular measurements for an unobserved modality and instead factorize the similarities between cells directly. UINMF, COBOLT and MIDAS previously handled paired and unpaired measurements through nonnegative matrix factorization of shared and unshared features, a multimodal variational autoencoder, and a probabilistic latent-variable model, each building a common representation at the feature or measurement level. Across the original 144 scMultiSim datasets, taking medians over four simulation replicates per fixed configuration, MCRF's median NMI was higher than UINMF in 22 of 36 configurations and higher than MIDAS in 28; the corresponding counts for higher ARI were seven and 15.

MCRF treats the cosine similarities between cells within each modality as the observed object and reconstructs them with shared cell factors and modality-specific matrices. RNA counts are normalized to 10,000 per cell and log1p transformed, ATAC counts are log1p transformed, and each cell vector is scaled to unit length so inner products give cosine similarities, after which six relation blocks are fitted jointly. Cells in the paired batch share one set of cell factors in both the RNA and ATAC reconstructions, each modality has its own matrix describing how factors produce similarity, each relation block carries an intercept, and the loss normalizes each block against its mean baseline before summing.

On the larger simulated datasets MCRF's absolute scores were generally higher than on the original collection, while its relative ranking fell. In the original collection MCRF combined higher median NMI with a smaller interquartile range in some configurations, whereas in the larger collection its median ARI was below both UINMF and MIDAS in all 24 configurations with four replicates. In discrete populations with 2,000 cells, 2,000 genes and CIF variability 0.1, MCRF's median ARI was 0.444 with an interquartile range of 0.001, against 0.683 and 0.079 for UINMF and 0.667 and 0.058 for MIDAS; its median NMI in the same configuration was 0.744, below UINMF at 0.800 and MIDAS at 0.831.

The evaluation is limited to simulated RNA and ATAC data generated by scMultiSim: 144 original datasets and 102 larger datasets, in a three-batch mosaic design with one batch containing both RNA and ATAC, one ATAC only, and one RNA only. ARI and NMI measure agreement with simulated population labels and do not separately assess batch removal, recovery of regulatory relationships, or preservation of other biological variation. The method's value on experimental data still needs to be established.

MCRF's absolute scores were generally higher in the larger collection than in the original one, yet UINMF and MIDAS had higher median ARI in all 24 configurations with four replicates, so absolute improvement and a lower relative ranking coexist. The two collections differ in cell counts, feature counts and other simulation settings at the same time, so the effect of increasing cell counts cannot be isolated. Raising CIF variability from 0.1 to 0.5 lowered MCRF's median NMI in 17 of 18 original comparisons and all 12 larger comparisons. The benchmark compares complete analysis procedures, and the methods differ in feature processing, latent dimensions, neighborhood construction and clustering algorithms; MCRF, UINMF and MIDAS all use reference labels to select the resolution with the highest NMI, so the scores describe label agreement after that selection. The interquartile ranges across simulation replicates describe variation between separately simulated datasets, not repeated initialization on the same dataset.