A Generalizable Transformer Framework for Gene Regulatory Network Inference from Single-Cell Transcriptomes

초록

Gene regulatory network (GRN) inference has advanced substantially through supervised and unsupervised learning approaches; however, many existing methods require extensive computational resources and exhibit limited generalizability across biological contexts. In addition, they do not fully exploit the growing availability of large-scale transcriptomic and regulatory data. Here, we present FTGRN (Foundation Transformer for Gene Regulatory Networks), a universal framework for GRN inference based on a pretrain-finetune paradigm. FTGRN integrates gene embeddings derived from Generative Pre-trained Transformer-4 (GPT-4) with publicly available chromatin immunoprecipitation sequencing (ChIP-seq) data to construct a regulatory knowledge base for pretraining a Transformer-based graph neural network. The pretrained model is subsequently fine-tuned using single-cell RNA sequencing (scRNA-seq) data to infer context-specific regulatory networks. Leveraging its pretrained foundation, FTGRN enables near real-time GRN inference, generating networks for 2,000 genes in under 30 seconds, substantially outperforming state-of-the-art methods in both speed and predictive accuracy. Application to amino acid-starved mouse embryonic fibroblasts demonstrated that FTGRN accurately reconstructs stress-response GRNs and identifies key regulators, including C/EBP gamma, c-Jun, DDIT4, and c-Fos. Collectively, FTGRN provides a scalable, adaptable, and interpretable framework for GRN inference in single-cell genomics.

키워드