scFoundation (BioMap)

scFoundation is a single-cell foundation model from China's BioMap that uses masked expression modelling over the full vocabulary of about 19k human genes, with an asymmetric encoder-decoder so the expensive encoder sees only expressed genes. The Nature Methods 2024 paper describes a 100-million-parameter model trained on 50 million cells and applies it to read-depth enhancement, drug response prediction and perturbation prediction. It is aimed at computational groups who want whole-transcriptome modelling rather than a truncated gene list, including for cancer cell line drug response. The design is compute-heavy, which limits who can fine-tune or run it at scale, and like its peers its perturbation predictions await independent benchmarking. For a newcomer: scFoundation models every gene in a cell at once and tries to predict how cells will respond to drugs.