Learning representations of single-cell genomics data is challenging due to the nonlinear and often multi-modal nature of the data on one hand and the need for interpretable representations on the other hand. Existing approaches tend to focus either …