From Closed-Set Annotation to In-Context Discovery: Biologically Grounded LLMs for Cellular Profiling

Tumour heterogeneity reflects both cellular identity (lineage) and state (active transcriptional programs), yet computational methods generally analyse these axes separately. We introduce CellScribe, a generative framework that conditions a pretrained language model on frozen single-cell foundation model embeddings through AdaLN adapters, enabling one shared model to generate an ontology- grounded cell type or a quantified meta-program profile. Supervised multitask fine- tuning produces fine-grained typing across hundreds of cell types and outperforms generative and retrieval baselines on meta-program identification under study-holdout evaluation. We further train the model with reinforcement learning, scoring its generations against cell ontology distance and meta-program agreement, so that they align with biological knowledge rather than token agreement. To identify new cell types not present in the supervised training data, we post-train the model to read several labelled example cells in a single context, each conditioning its own text span, and infer the type of a query cell through in-context learning. CellScribe thus unifies cell identity and state in one generative interface, grounds both in biological structure, and identifies cell types outside the supervised vocabulary.