ReCellTy

Annotating cell types in single-cell RNA-seq is a two-step act of expertise: pull the marker genes that distinguish a cluster, then combine them with domain knowledge to name the cell. Large language models handle the second step passably, but they do the first from parametric memory, which is where they hallucinate marker-cell associations that read fluently and appear in no curated database. Fine-tuning an LLM on annotation data trades one problem for another, since continued fine-tuning erodes previously learned knowledge.

ReCellTy, from the School of Control Science and Engineering at Shandong University with Qilu Hospital and Johns Hopkins University, takes the retrieval route instead. It reconstructs the CellMarker 2.0 database into a Neo4j knowledge graph designed for annotation, then drives a general-purpose LLM through a fixed five-task workflow that queries that graph before it commits to a name. Nothing is trained: the LLM is swappable (the paper evaluates GPT-4o, GPT-4o-mini, DeepSeek-chat, and Claude 3.7 Sonnet), and the contribution is the graph plus the orchestration around it. The name stands for "Reconstructing Cell Types" — rebuilding a cell name from its parts rather than retrieving it whole.

#Key Features

• Annotation-oriented graph reconstruction: Cell names are decomposed into naming features and broad cell types, so that "CD4+ central memory T cell" becomes queryable components linked to their markers rather than one opaque string.
• Dual-retrieval augmentation: For a given marker set the system retrieves broad cell types and functional feature descriptions in parallel, giving the downstream agents a category-level prediction and a structured feature table as separate evidence streams.
• Multi-task agent workflow: Five tasks — cell type query, cell type selection, feature query, feature selection, and final annotation — mirror the sequence a human annotator follows, and each intermediate decision is inspectable.
• Model-agnostic backend: Any general-purpose LLM can sit behind the workflow, and the gains hold across all four tested models rather than depending on one vendor's checkpoint.
• Pipeline integration: A Python package handles top differentially expressed genes and interoperates with R workflows such as Seurat via rpy2 and reticulate, alongside a UI exposing retrieval progress and agent decisions.

#Technical Details

The reconstruction pipeline processes over 78,000 raw human entries from CellMarker 2.0, using an LLM to parse feature and function descriptions out of cell names, identify broad cell type classifications, and build explicit feature-marker mappings. The result is 1,528 naming features and 61,049 feature-gene-cell-type associations, loaded into Neo4j as 18,850 entity nodes and 48,944 relationships across seven node types (including Marker and FeatureFunction) and seven relation types. Retrieval runs through LangChain's GraphCypherQAChain: the LLM receives a schema description of the graph, generates Cypher from a natural-language query with annotation-specific prompt templates and worked query examples, executes it, and reasons over the returned rows.

Evaluation used 11 tissue types from the Azimuth reference dataset, with five independent question-answering iterations per differential-gene row and the modal result taken as the answer. Human evaluation scores annotations on a 0–1.5 scale (1.5 when the prediction is more specific than the manual label, 1.0 exact, 0.5 same major type, 0 mismatch); semantic evaluation embeds both annotations with text-embedding-3-small and bins cosine similarity. ReCellTy beat both general-purpose prompting and the CellMarker 2.0 web tool on all four backend models under both metrics, with an average human-eval gain of 0.18 and a maximum of 0.21 on DeepSeek-chat, and a 3.8% average semantic-similarity improvement. The most striking result is compression of the model gap: GPT-4o-mini rises from 0.50 to 0.67, overtaking bare GPT-4o (0.62), DeepSeek-chat (0.55), and Claude 3.7 (0.61).

#Applications

ReCellTy suits labs that want automated first-pass annotation traceable back to a curated source. The graph is fixed and the workflow is unchanged between tissues, so it applies to new datasets without per-tissue fitting, and the exposed intermediate steps — retrieved broad cell type, selected features — let an analyst audit a call instead of accepting it. Its higher intra-group annotation diversity also makes it more willing to name rare subtypes that generic prompting flattens into a common label.

#Impact

ReCellTy is a clean demonstration that grounding matters more than model scale for this task: a small model with a good graph beat larger models without one. It sits alongside mLLMCelltype, which reaches the same goal through multi-LLM consensus voting rather than domain retrieval, and the two illustrate the dominant strategies for making LLM annotation trustworthy. The caveats are real: gains over general-purpose prompting are modest in absolute terms, evaluation is confined to one benchmark, and the system depends on third-party commercial APIs, so results are not fully reproducible over time. The work remains a preprint. The graph construction code, demo, and UI are on GitHub, with the authors noting that some components await manuscript acceptance, and no license file is present on the repository.