Fig. 5

Benchmark evaluating the classification accuracy for closely related immune cell types in nine datasets using the detailed cell annotations (level 2). The benchmark was performed in a ninefold cross-validation scheme where one dataset was used for training and the other eight for testing. All metrics were ordered by the tools’ mean accuracy. (*) indicates results where intermediate/ambiguous classifications were accepted. scibet-rej represents SciBet with a confidence threshold of 0.4. Blank areas indicate comparisons that could not be performed with the respective tool. a The prediction accuracy in each training set and test set for all tools. Panels show the sensitivity (b), specificity (d). The shown values (points) represent the mean sensitivity and specificity, respectively, across all classified cell types across the five evaluated datasets per iteration. Panel (c) shows the proportion of cells not present in the training data, that were correctly not-classified