CellDecipher: Automated Cell Phenotyping in Multiplexed Imaging with LLM derived priors

Highly multiplexed antibody-based spatial proteomics reveals the spatial organisation and localisation of distinct cell populations at a single cell resolution, and has become a key technology for dissecting tumour microenvironments and immune organisation in support of translational applications. A central step in the analysis of such data involves the inference of cell types and their states, i.e. cell phenotyping. Phenotyping is a complex task, usually requiring extensive manual expert intervention to derive patterns of marker abundances specific to each cell type or state of interest. The relationships between specific markers and the cell types differ by the selected panel, tissue and disease context. We propose CellDecipher, an automated framework that efficiently extracts and maps domain-specific prior knowledge from Large Language Models (LLMs) to enable automated and interpretable cell phenotyping. Rather than querying the model per cell or per cluster of cells with similar profiles, we extract relevant knowledge into a discrete, human-auditable marker-by-cell-type matrix that is then applied to obtain interpretable priors to cold start solving the phenotyping problem as a classification task. We further extend this approach by introducing an optional human-in-the-loop step via a dedicated visualisation tool, allowing users to efficiently target and revise initial low-confidence or ambiguous cell assignments to provide improved, high-confidence consensus labels for training a final classifier. Across four expert-annotated datasets, CellDecipher, even without the optional human-in-the-loop, outperforms existing methods, whether those methods use expert encoded decision matrices or those made by related LLM-based approaches automatically. We demonstrate that CellDecipher outperforms an approach to directly prompting assignments of individual clusters and exceeds the macro-F1 performance of related approaches, indicating better recovery of rare populations.