Fig. 1

Overview of the joint embedding in JVis. Metrics d (left) and d^′ (right) measure the dissimilarity of different cellular phenotypes of individual cells, such as the expression of surface proteins (left) and mRNA (right). t-SNE and UMAP learn a low-dimensional embedding of cells that preserves the distribution of similarities that are quantified based on d or d^′ alone, which renders certain cell types indistinguishable to either modality. In this example, blue and red cells cannot be distinguished based on their measured surface proteins, and green and black cells overlap in transcriptomic space. In JVis we generalize t-SNE and UMAP to learn a joint embedding that preserves similarities in all modalities at the same time. We integrate d and d^′ in a convex combination of KL divergences (j-SNE) or cross entropies (j-UMAP) between corresponding similarities in low and high-dimensional space. An arrangement of cells that minimizes this convex combination with simultaneously learned weights takes into account similarities and differences in both mRNA and surface protein expression to more accurately represent cellular identity (middle)