CohortHijack: Robustness of Single Cell Annotation to Companion Cell Removal

CohortHijack targets single-cell annotation tools that refine initial cell labels using nearby cells or cluster-level voting. The audit preserves the target expression profile, base prediction, and trained model while removing selected non-target cells, revealing potential vulnerabilities in refinement mechanisms. The study compares random and structured removal methods, and likely benchmarks multiple annotation tools. This work matters because it highlights a previously underappreciated attack surface in single-cell analysis pipelines, where small changes in the cell cohort could alter final labels. The findings could lead to more robust annotation methods in computational biology.