Fig. 2: Model outputs and downstream tasks.

a The cell embeddings are derived either by aggregating gene tokens within each cell or by appending and training special tokens (cell tokens, often classification tokens) that represent individual cells. b The gene embeddings are produced by pooling gene-specific embeddings from all input cells. In addition, attention scores from selected layers and heads capture the relationships among genes. c The models generate feature profiles that represent the distribution of cells across different conditions or states. a–c explain initial outputs generated by feeding inputs to scFMs. d Various downstream tasks with scFMs. Each task is colored with types of model-generated output utilized for the task.