scGPT

scGPT is a generative transformer for single-cell data that tokenises genes and bins expression values, borrowing the GPT recipe from language modelling. The Nature Methods 2024 paper pretrained it on 33M cells and showed fine-tuning for cell-type annotation, batch integration, perturbation response prediction and gene regulatory network inference. It is used by computational biologists as a general-purpose starting point that can be adapted to a new dataset with limited labels, including tumour microenvironment atlases. The main caveat is that its perturbation prediction is only modestly above simple baselines, a finding echoed across single-cell foundation models, so its strengths are in annotation and integration rather than in forecasting drug effects. For a newcomer: scGPT is a GPT-style model for cells that is good at naming cell types and merging datasets.