Skip to main content
Fig. 3 | Genome Biology

Fig. 3

From: Schema: metric learning enables interpretable synthesis of heterogeneous single-cell modalities

Fig. 3

Incorporating temporal metadata into UMAP visualizations of aging neurons captures developmental changes. UMAP visualization of RNA-seq profiles of D. melanogaster neurons at 0, 1, 3, 6, 9, 15, 30, and 50 days after birth, representing the full range of a typical D. melanogaster lifespan. The transcriptomic data (primary modality) was transformed to a limited extent using Schema by correlating it with the temporal metadata (secondary modality) associated with each cell. a UMAP visualization of the original transcriptomic data. b–d Visualizations of transformed data with increasing levels of distortion. As the value of the minimum correlation constraint s approaches 1, the distortion of the original data is progressively limited. Decreasing s results in a UMAP structure that increasingly reflects an age-related trajectory. e Feature selection interpretation of Schema’s transformation. In synthesizing the two modalities, Schema up-weights genes (top 15 shown here) that are differentially active at the start or end of the time-course. For clarity, the set of genes has been reordered by the difference in their early and late-stage expression

Back to article page