scROSHI - single cell robust supervised hierarchical identification of cell types

Identification of cell types based on expression profiles is a pillar of single cell RNAseq analysis. Many existing methods utilize machine learning to identify predictive features from annotated training data. In practice, this procedure can lead to overfitting and inferior performance when applied to previously unseen data. In addition, this procedure always requires annotated data for building a prediction model, which are not available in early stage studies. Finally, in the most common setup all available cell types are classified simultaneously, which can be challenging if they are not very distinct. To address these challenges, we present scROSHI, a cell type classification method that utilizes cell type specific gene sets derived from literature or previous experiments. scROSHI performs robust hypothesis testing that does not require training or the existence of annotated data. Furthermore, by respecting the hierarchical nature of cell type relationships and assigning cells consecutively to more specialized identities, excellent prediction performance can be achieved. scROSHI outperforms competing methods when training data are limited or the diversity between experiments is large.