Describir: Learning interpretable cellular and gene signature embeddings from single-cell transcriptomic data