scGPT

scGPT (Nature Methods, 2024) is a transformer foundation model for single-cell multi-omics. It was pretrained on 33M human cells with a generative gene-expression-prediction objective, then fine-tuned per task: cell-type annotation, batch integration, perturbation prediction, multi-omic integration, and gene-network inference.

The code and pretrained weights are fully open (MIT-licensed on GitHub) with tutorial notebooks covering each downstream task, and organ-specific variants (brain, lung, heart, blood) were also released. It is the most widely used starting point for single-cell foundation-model work, though zero-shot performance on unseen datasets remains limited — fine-tuning is expected.

Evaluated on cell-type annotation, multi-batch integration, perturbation prediction, and gene-network inference in the Nature Methods 2024 paper; follow-up benchmarks compare it against Geneformer, scFoundation, and UCE — see the paper and the scFMs_Dynamic benchmark repo for numbers.