Fig. 5: Benchmarking of multiplet detection on the Ileum-1 dataset.
From: A unified model-based framework for doublet or multiplet detection in single-cell multiomics data
a–d UMAP plots displaying the comparison between multiplet predictions and the ground truth on the Ileum-1 dataset. The multiplet detection methods shown are COMPOSITE (DOGMA) (a), COMPOSITE (RNA) (b), scDblFinder (RNA) (c), and DoubletFinder (RNA) (d). True positive (Prediction \(\cap\) Ground truth), false positive (Prediction \(\notin\) Ground truth), and false negative (Ground truth \(\notin\) Prediction) predictions for multiplets are highlighted with green, red, and dark blue, respectively. Comparing the results from a–d, the circled cluster shows the most prominent difference among the prediction results from different methods. In the circled cluster, the three prediction methods based on RNA data resulted in many false positives (b–d), while the false positive rate for COMPOSITE (DOGMA) remained low (a). e Manual cell type annotation based on ADT data after removing ground truth multiplets. The clustering and UMAP visualization were generated from weighted nearest neighbors using all three modalities of data39. Source data are provided as a Source Data file.
