Clifti-GPT: Privacy-preserving federated fine-tuning and transferable inference of foundation models on clinical single-cell data

Foundation models have demonstrated immense value for scRNA-seq analysis, but their fine-tuning or inference on heterogeneous, privacy-sensitive clinical cohorts is governed by strict data protection policies, which often prohibit centralization. We introduce clifti-GPT, a privacy-preserving federated solution that leverages secure multiparty computation to enable collaborative model training and transferable inference of local statistics in zero-shot applications across decentralized scRNA-seq clinical repositories, without sharing patient data or clinical-level statistics or models. Built upon the scGPT foundation model, clifti-GPT achieves performance within 4% of centralized baselines in accuracy, precision, recall, and macro-F1 for cell type classification and reference mapping across six datasets. Furthermore, it demonstrates high communication efficiency, reaching 99% of centralized performance in fewer than two rounds, and scales robustly to 30 clients with less than 2% accuracy loss. Thus, clifti-GPT makes it feasible to fine-tune and apply single-cell foundation models across distributed clinical datasets under real-world privacy and governance constraints.