Single-cell sequencing now measures several omics layers in the same cell—transcriptome, epigenome, proteome and beyond—and integrating them offers a far fuller view of cellular heterogeneity than any single layer. Yet integration remains difficult: the data are sparse, noisy and heterogeneous, the modalities differ in dimensionality, and, most importantly, most methods do not preserve the true regulatory relationships among molecular features. Matrix factorization approaches such as MOFA assume essentially linear cross-omics mappings; manifold alignment approaches such as MACHER need strong prior assumptions and are costly. Deep learning methods such as totalVI, Cobolt, DCCA and scMIC, and representation-learning frameworks such as scDRMAE and SIMBA, but these mainly learn cell-level embeddings and seldom extract feature embeddings, which limits downstream regulatory analysis.
Recently, a group led by Prof. Hongwei Li at China University of Geosciences, published a research article, “scMOG: A graph neural network method for regulatory relationship‐preserving single‐cell multi‐omics integration,” in
Quantitative Biology. For each omics layer, scMOG builds a heterogeneous graph that links cell nodes directly to feature nodes—genes, chromatin regions or proteins—and applies LightGCN encoders, so that every graph convolution passes messages between features carrying prior regulatory or interaction links. Zero-inflated negative binomial decoders reconstruct the sparse counts; a contrastive learning module built on mutual nearest neighbour pairs and a triplet loss sharpens cell representations; and an omics alignment module holds the same cell close across omics. Eight paired mouse and human datasets and nine competing methods were used for evaluation.
Figure 1 illustrates the workflow: input matrices are preprocessed separately and then coupled through the graph autoencoder, the contrastive learning module and the omics
alignment module, which learn cell and feature representations for each layer without supervision, while clustering, differential expression and evaluation metrics assess the
outcome. On the Chen dataset (22 cell types), scMOG reached ARI 0.71, NMI 0.73 and ACC 0.81, separating populations such as Clau and E5Tshz2 more cleanly than existing tools, whereas scGAE struggled on the same data. On the Spector dataset (16 cell types) it led every method with ACC 0.69, NMI 0.71 and ARI 0.58. Beyond clustering, its gene–peak similarity matrices agree with annotation-derived associations better than direct preprocessing and beat SIMBA on three of five epigenomic datasets, and its protein embeddings recovered interacting pairs SIMBA missed, including CD28–SELL, ITGAL–TNFRSF9 and Ccr4–Ms4a1. Ablation experiments confirmed that the LightGCN encoder and both auxiliary modules are essential.
DOI:
10.1002/qub2.70041