Single-cell datasets are increasingly costly to store, audit, and reuse for model training. While dimensionality reduction and dataset distillation can alleviate this burden, conventional methods often produce synthetic expression profiles that cannot be traced back to actual cells. We define traceable single-cell data distillation as retaining original cell identifiers and gene symbols under fixed cell and gene budgets. The resulting training subset remains connected to measured counts, labels, and assay metadata, allowing unexpected predictions to be checked against their source data.
We propose two real-cell selectors. Fixed-CF uses static characteristic-function matching, while Minmax-CF addresses an entropy-regularized discrete min-max problem that emphasizes poorly preserved directions and includes only observed cells.
Across donor, technology, and perturbation-level shifts in three datasets, Minmax-CF retains 96.52% of Full balanced accuracy on MS, approximately matches Full on average on hPancreas with a median $2.55\times$ GPU speedup in the all-gene setting, and achieves the lowest pathway error among compressed methods on Norman.
Performance remains weaker for rare states, certain technology shifts, unseen perturbation components, and settings where fidelity is weakly associated with downstream utility. Since the selected IDs refer to measured cells, these cases can be investigated by inspecting the corresponding training support, labels, and assay metadata. Minmax-CF consistently reduces worst-direction discrepancy, while downstream utility and cost vary across datasets and tasks.
Blogger's Review: The Minmax-CF method proposed in this paper significantly enhances training efficiency while ensuring data traceability, showcasing the potential of dataset distillation in the field of bioinformatics. By addressing cell and gene budget issues, the researchers provide new insights and tools for future single-cell data analysis.