Overview

• The GREmLN preprint, posted July 10 on bioRxiv, introduces a graph-based architecture that captures long-range gene–gene relationships in regulatory networks.
• It is trained on more than 11 million single-cell RNA-seq profiles from CZI’s CZ CELLxGENE database, predominantly sourced from healthy human donors.
• In benchmarks, GREmLN outperformed leading scRNA foundation models—Geneformer, scGPT and scFoundation—in cell-type annotation and network structure tasks using fewer parameters.
• The model joins CZI’s open virtual cell platform alongside TranscriptFormer, offering global researchers access to its code and pretrained weights.
• Experimental validation is underway using a new genetic perturbation dataset targeting druggable cancer genes as teams prepare to expand GREmLN beyond transcription to include other regulatory layers.