Empirical Evaluation of Single-Cell Foundation Models for Predicting Cancer Outcomes

Foundation models pretrained on large-scale single-cell RNA sequencing data present a promising opportunity to advance translational cancer research. However, their utility in clinically relevant, patient-level applications of single-cell analysis remains underexplored. Here, we systematically evaluated nine emerging single-cell foundation models (scFMs) and three alternative baseline approaches across six cancer-specific tasks, ranging from subtype classification to treatment response prediction. We assessed model performance under zero-shot, continual training, and fine-tuning conditions, conducting 1,170 supervised and 130 unsupervised experiments. Our findings revealed that while current scFMs excel in certain analysis tasks, such as tumor microenvironment cell annotation, they had limited advantages in predicting clinical and biological outcomes of cancer patients compared to simpler baseline models. These insights highlight the critical role of evaluation on biologically and clinically relevant tasks in the responsible and impactful application of scFMs in precision oncology. They also emphasize the necessity of further methodological innovation and expanded cancer single-cell cohorts for the future development of scFMs.