Abstract
<title>Abstract</title> <p>A classifier can label single cells accurately while learning study-specific or person-specific signals. We tested how much this concern matters for a narrow, common task: annotation of abundant pancreatic cell types within one single-cell RNA-sequencing study. We reanalyzed 8,569 human pancreatic cells from four donors and retained six labels with at least 20 cells in every donor, giving 7,068 cells. Counts were library-size normalized and log transformed. In each training fold, the 2,000 most variable genes were selected before fitting either a class-balanced multinomial logistic regression or a cosine-centroid classifier. We compared five-fold random cell splitting, which places cells from every donor in both sets, with leave-one-donor-out evaluation. Logistic regression reached mean macro-F1 0.9898 under random splitting and 0.9853 with a donor held out, a difference of 0.0045. The centroid classifier gave 0.9853 and 0.9831, respectively. Across held-out donors, logistic macro-F1 ranged from 0.9761 to 0.9913. Results were stable when the feature count varied from 250 to 4,000 genes. Thus, for abundant, canonical pancreas classes measured by one protocol, most cell-type signal generalized across people and random cell splitting caused only modest optimism. This finding should not be extended to rare populations, disease states, cross-study transfer, or mechanism. The analysis illustrates a broader point: claims about knowledge extracted from single-cell data must state the population and technical shift across which that knowledge was tested.</p>