Evaluation of some aspects in supervised cell type identification for single-cell RNA-seq: classifier, feature selection, and reference construction

Table 2 Before and after “purification”

	Original performance	After distance-based purification	After probability-based purification
Human PBMC lupus: one individual predicts another individual	0.797	0.748	0.799
Human PBMC lupus: one batch predicts another batch	0.924	0.920	0.921
Human PBMC lupus: one status predicts another status	0.931	0.932	0.934
Mouse brain FC: one subject predicts another subject (sub-cell types)	0.783	0.813	0.802

Performance comparisons before and after purifications on reference dataset. Here, we only demonstrate the performance of overall accuracy. There are in total four experiments: (1) “Human PBMC lupus” one individual predicts another individual: uses one lupus patient from batch 1 to predict another from the same batch under the same condition; (2) “Human PBMC lupus” one batch predicts another batch: uses samples from batch 1 to predict samples from batch 2 under the same condition; (3) “Human PBMC lupus” one status predicts another status: uses lupus samples from batch 2 to predict IFN-β stimulated samples from the same batch; and (4) “Mouse brain FC” one subject predicts another subject: uses one mouse subject to predict sub-cell types of another subject from the same region and the same dataset. The bold data indicates a performance improvement

ISSN: 1474-760X