The paper introduces CellFM, a novel large-scale foundation model specifically designed for single-cell transcriptomics analysis, built upon a dataset of 100 million human cells. This advanced model, featuring 800 million parameters, aims to overcome the limitations of existing single-cell models by offering enhanced capabilities in tasks such as cell type annotation, perturbation prediction, and gene function prediction. CellFM leverages a modified RetNet architecture, dubbed ERetNet, for improved training efficiency and performance, alongside a Low-Rank Adaptive (LoRA) module to optimize fine-tuning. Extensive experiments demonstrate CellFM's superior accuracy and robustness across various biological applications, showcasing its potential to significantly advance the understanding of cellular states and gene interactions.

References:

• Zeng Y, Xie J, Shangguan N, et al. CellFM: a large-scale foundation model pre-trained on transcriptomics of 100 million human cells[J]. Nature Communications, 2025, 16(1): 4679.