Fig. 1: Overview of scMODAL.

a scMODAL takes single-cell feature matrices from different modalities, together with feature links as input. b scMODAL utilizes generative adversarial learning to mix the distributions of cell embeddings from different datasets. To find correct correspondence between modalities as well as preserve biological variation within each modality, regularizations to narrow the distance between anchors based on prior information and preserve geometric representation of cells are applied in the training process of scMODAL. c scMODAL outputs integrated cell representations for further analyses, and the composition of trained networks enables imputation of features and inference of feature relationship across single-cell modalities. The results can also be used for multiple downstream analyses, including label transfer for revealing cell identities and cell-cell communication inference using imputed features.