Fig. 1

Overview of MinNet. A Model receives two modalities’ data as input. High-throughput omics data will go through an independent fully connected layer to be projected into a lower dimensional space. This representation space should be able to mix different modalities and separate cell types well. To achieve this, cell type classification loss and Siamese contrastive loss are used during the training process. B To make the mixing resolution at single-cell level rather than cell-type level, we applied a KNN graph-based Siamese loss with flexible margin value depending on cell pair graph distance. C In application, multiple omics data will be projected into this low-dimensional embedding space in which downstream analysis will be done, including cell alignment, label transfer, unsupervised clustering, and the designed cis-regulatory element-inferring pipeline