A benchmark of semi-supervised scRNA-seq integration methods in real-world scenarios.

Semi-supervised methods for single-cell RNA-seq integration promise improved batch correction and preservation of biological signal by leveraging cell-type labels. However, reported benefits and robustness of them towards imperfect cell type labels often come from overly idealized settings. Here we present, to our knowledge, the first systematic benchmark comparing leading semi-supervised methods with widely used unsupervised approaches across six diverse datasets under realistic conditions. Beyond randomly missing or erroneous labels, we examine four additional scenarios (boundary-mixed labels, batch-specific annotations, auto-generated labels and varied-granularity labels) and evaluate performance using nine established metrics. We find that although semi-supervised methods can provide benefits under perfect annotations, their robustness often degrades substantially under realistic imperfections. Only scANVI and ssSTACAS maintain stable but modest improvements over their unsupervised counterparts, and none consistently outperform the strongest unsupervised approach. These results indicate that current semi-supervised strategies offer limited practical advantage when label quality is modest uncertain.