Two-step workflow of creating a foundation model for single-cell genomics: training an LLM on >10,000,000 transcriptome sequencing observations (CellXGene portal) in a self-supervised fashion (i.e., without predicting target phenotypes or classifications), the thus pre-trained LLM can then be leveraged to great effect via parameter adaptation pipelines (fine-tuning) for increased performance in specific application tasks on smaller, unseen snRNA-seq datasets (few-shot learning).
