Benchmarking single-cell foundation models in a zero-shot setting

Single-cell foundation models are trained on millions of cells to learn general-purpose representations that can be transferred to downstream tasks. In this preprint, the authors systematically compare scGPT, SCimilarity, UCE, and Transcriptformer against conventional methods without task-specific fine-tuning, using a zero-shot evaluation protocol. The benchmark covers multiple tasks to determine whether these models provide meaningful performance gains over established tools. Results from this study are directly relevant for researchers weighing the adoption of foundation models in single-cell genomics workflows. This work helps quantify the actual benefits of these models, moving beyond anecdotal claims and providing actionable guidance for the community.