From Closed-Set Annotation to In-Context Discovery: Biologically Grounded LLMs for Cellular Profiling
- 1InstaDeep
Abstract
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.
Related
- Biologically grounded cell profiling across microscopy modalities
- Biologically grounded locality priors close the data gap for vision transformers in neural prediction
- Reconstruction of Cellular Trajectories with Biologically Interpretable States
- Integrating multi-omics with machine learning: from genomic variant functional annotation to pharmacogenomics and biomarker discovery
- A low-annotation-budget PubMedBERT classifier for chondrogenesis regulator discovery via active learning
BibTeX
@misc{gelard2026from,
title = {{From Closed-Set Annotation to In-Context Discovery: Biologically Grounded LLMs for Cellular Profiling}},
author = {M. Gelard and H. Benkirane and T. Pierrot and P.-H. Cournede and B. P. de Almeida},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.07.757352},
url = {https://www.biorxiv.org/content/10.64898/2026.10.07.757352v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - From Closed-Set Annotation to In-Context Discovery: Biologically Grounded LLMs for Cellular Profiling
AU - M. Gelard
AU - H. Benkirane
AU - T. Pierrot
AU - P.-H. Cournede
AU - B. P. de Almeida
PY - 2026
DA - 2026/10/07
PB - bioRxiv
DO - 10.64898/2026.10.07.757352
UR - https://www.biorxiv.org/content/10.64898/2026.10.07.757352v1
AB - 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.
ER -