Benchmarking of automated cancer cell annotation methods for scRNA-seq data reveals Consensus annotation as the preferred method

Targeted cancer therapies have shown therapeutic advantages due to tumor-specific drug activity. Single-cell RNA-sequencing has been widely used in cancer studies to define different cellular identities. However, accurate identification of tumor vs other normal cells is essential to define novel tumor-specific targets. Recent methods have been developed to perform the task of tumor cell annotation, which can be divided into two categories: CNV-based methods, which use transcriptome measurements to infer copy number variations and identify cells with alterations as tumor, and Reference-based methods which train a classifier using previously annotated tumor data and use it to annotate new datasets. We benchmarked the state-of-the-art method of each category, SCEVAN and scATOMIC, respectively, together with Consensus annotation method where a cell is considered tumor if both methods agree on that. Across 20 cancer datasets spanning 9 cancer types with a total of 379 samples, the Consensus annotation outperformed other methods in terms of precision score. SCEVAN works well when clear CNVs are detected, otherwise cells are randomly split between normal and tumor producing many false positives. While scATOMIC efficiently detects all normal cells except normal epithelial cells, where all epithelial cells (normal and malignant) are considered cancerous. The Consensus annotation method overcomes the limitations of both methods, being able to detect normal epithelial cells together with other normal cell types as non-tumor, also annotating malignant epithelial cells as tumor even with weak CNVs. This produces overall higher precision with the least number of false positives, leading to confident tumor-specific potential therapeutic targets. Implementation is available in the MACE R package through GitLab (https://gitlab.com/genmab-public/mace/).