Single-cell genomics is enhancing our understanding of life by making sense of novel data from next-generation sequencing protocols using algorithmic advances in machine learning. Consequently, efforts like the Human Cell Atlas are creating large harmonized datasets that capture cellular heterogeneity across individuals and tissues, so called integrated single-cell reference-atlases. These atlases can be used to automatically annotate new datasets. The first project of my PhD uses Bayesian Machine learning methods to provide uncertainty estimates alongside cell type predictions. This not only enables more reliable annotation, but also discovery of new biology by using epistemic uncertainty. Secondly, with the increasing size of single-cell datasets, individual-level analyses become feasible. Applications are discovery of transcription regulating genes (eQTLs) and disease risk scores. These challenges require methods from Deep Generative Modelling in general and multi-instance learning in particular.