scGPT, a Single-Cell Foundation Model Trained on 33 Million Cells, Is Released
Researchers at the University of Toronto handled cell type classification and perturbation response prediction in one model trained on single-cell transcriptomes as if they were 'sentences.'
Single-cell RNA sequencing (scRNA-seq) data have piled up explosively, but for each experiment researchers had to pick a new analysis tool and correct batch effects separately. Researchers at the University of Toronto and the Vector Institute in Canada reported in Nature Methods scGPT, a foundation model that carries the language-model approach over to cell data.
Pretrained With 'Sentence = Cell, Word = Gene'
The researchers pretrained a generative transformer on 33 million cells from public single-cell data. They applied directly to the design the analogy that, just as a sentence is made of words, a cell is defined by a combination of gene expression. When the resulting model was fine-tuned by transfer learning, it could be used for different tasks, including cell type annotation, integration of multiple batches, multi-omics integration, prediction of gene perturbation responses, and inference of gene regulatory networks.
The significance is that the axis of analysis pipelines has shifted from 'a dedicated algorithm for each task' to 'a model trained once and attached to many tasks.' Even for tissues and diseases with little data, pretrained knowledge can be borrowed.
Limits
Performance depends on the quality of the data used for fine-tuning and on tissue and ancestry biases in the pretraining data. The regulatory networks or perturbation predictions the model produces are hypotheses and must be confirmed by experiments.
- Generative transformer foundation model pretrained on 33 million cells of public data
- Applied across cell type annotation, batch integration, multi-omics integration, perturbation prediction and regulatory network inference
- Shift from task-specific tools to fine-tuning a pretrained model
Limitations Tissue and population biases in the pretraining data are reflected in performance, and the network and perturbation predictions the model produces are hypotheses that need experimental validation.
Original research
- Cui H, Wang C, Maan H, … Wang B · University of Toronto / Vector Institute (Canada). scGPT: toward building a foundation model for single-cell multi-omics using generative AI. Nature Methods. DOI: 10.1038/s41592-024-02201-0
Copyright in the original research belongs to its publisher and authors.