New 'CohortHijack' method reveals vulnerabilities in single-cell annotation tools

Researchers have developed a new method called CohortHijack to test the robustness of single-cell annotation tools. This technique involves removing specific non-target cells from a dataset to see if it alters the annotation of the remaining target cells. Experiments on PBMC3K and Paul15 datasets using logistic regression and linear SVM classifiers showed that structured removal of companion cells could significantly change target cell labels while preserving the original model and target cell expression profiles. The study highlights that the composition of the query cohort is a potential vulnerability in single-cell annotation processes. AI

IMPACT This research highlights a potential attack surface in single-cell annotation, suggesting a need for more robust methods in biological data analysis.

RANK_REASON The cluster contains a research paper detailing a new method for auditing the robustness of single-cell annotation tools. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• CellTypist
• CohortHijack
• Linear Svm
• logistic regression model
• Paul15
• PBMC3K

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →