Fig. 5: Mutual information-based disentanglement enables the discovery of batch-confounded cell types.

a “Frankencell” synthetic dataset generation and evaluation algorithm. Starting from either of two batches in the NEURIPS bone marrow dataset which exhibit different technical effects, reads from annotated cell clusters were mixed according to a construction plan to create a simulated differentiation trajectory interpolating between cell types. Datasets were composed of a trajectory constructed from both batches to represent known sources of biological and technical variation. By controlling terminal cell states present in each trajectory, we introduced batch-confounded cell types. We also varied the base cell similarity to measure method robustness. The trajectories were then integrated and evaluated against the construction plan using established trajectory comparison metrics. b Results from F1 branch score metric across all tests, colored by method. F1 branch score measures the similarity of predicted cell branch assignments to the construction plan. All branch assignments were calculated using MIRA pseudotime trajectory inference on the integrated trajectory. c Example UMAPs of CODAL and Harmony latent spaces, colored by true cell type and batch of origin. Cells from different batches were slightly offset for readability. Each example was taken from the “Medium” base cell similarity test. Source data are provided as a Source Data file.