Fig. 5.
While existing models struggle with over- or under-integration, the VAMP + CYC model achieves a good trade-off between batch correction and biological preservation. a We show the integration performance of individual models using the best hyperparameter settings, with different integration metrics in columns (larger values indicate better integration performance) and cross-system data cases in rows, showing the average performance (vertical line) of three runs (dots). The results of all hyperparameter settings are in Supplementary Figure S1. b UMAPs of representative runs for the best hyperparameter setting in the pancreatic mouse-human dataset. A legend for sample colors is not shown due to the large number of samples. UMAPs for the other datasets are shown in Supplementary Figure S5. c Schematic representation of the VAMP + CYC model in accordance with the explanation of the VampPrior and latent cycle-consistency loss from Fig. 1d