Consequences of training data composition for deep learning models in single-cell biology