Cross-Donor Generalization
Testing how evaluation design changes estimates of cell-type classifier performance on unseen donors.
Cell-type classifiers can appear more accurate when training and test sets contain cells from the same donors. With Jiabei Li and Alexis Battle, this project compares random cell-level splits with donor-held-out evaluation across multiple single-cell RNA-seq datasets. The central question is when cell-level testing overstates performance on unseen donors, and how that gap depends on donor and cell-type structure.
I developed reproducible benchmarking workflows to compare classifiers using highly variable genes, PCA, Harmony, and scVI representations. Analyses examine overall classification performance alongside donor-level variation and cell-type-specific errors, helping distinguish an evaluation-design effect from a representation or dataset effect. The goal is to clarify what a reported accuracy estimate says about generalization to new individuals.
Manuscript in preparation: Evaluation design shapes estimates of cross-donor generalization in single-cell cell-type classification — Xingyi Chen, Jiabei Li, and Alexis Battle.
