CellReasoner

Assigning a cell type to each cell is the step everything downstream of a single-cell experiment depends on, and it is fundamentally an act of reasoning: weigh which marker genes are present, which are conspicuously absent, and what the tissue context implies. Human experts do this by argument. Most computational annotators do not — they cluster, then match centroids to a reference, and any clustering error propagates to every cell in the cluster.

CellReasoner, from Dijun Chen's group at Nanjing University with collaborators at Zhejiang University and Inner Mongolia Minzu University, takes the reasoning framing literally. It is a 7B-parameter open-weights language model fine-tuned to read a single cell's expression profile, rendered as a "cell sentence" of ranked expressed genes, and emit both a marker-by-marker reasoning chain and a final cell-type label. Annotation happens per cell, not per cluster, so a mislabelled neighbourhood does not contaminate its members.

This is a distinct approach from prompting-based systems such as mLLMCelltype, which orchestrates a consensus among off-the-shelf general LLMs over cluster marker lists without training a new checkpoint. CellReasoner instead produces a purpose-trained model, and its central claim is about training efficiency: the authors argue that a 7B open model already encodes rich marker-to-cell-type knowledge, so the task is knowledge activation, not acquisition. Their CRAFT strategy activates it with only 380 chain-of-thought exemplars.

#Key Features

• Per-cell annotation: Predictions are made on individual cells rather than clusters, allowing fine-grained subtype resolution and local error correction instead of cluster-wide propagation.
• Compact CoT training: The three-stage CRAFT recipe — reasoning scaffold, knowledge infusion, then reasoning mode fusion to restore reasoning degraded by the second stage — uses only 380 curated chain-of-thought exemplars.
• Interpretable output: The model states which markers it used and why, including negative exclusions, producing an auditable rationale rather than a bare label.
• Consumer-hardware inference: The 7B checkpoint runs on a single NVIDIA RTX 4090, with inference time scaling roughly linearly in cell count.
• Few-shot adaptation across modalities: Trained only on scRNA-seq, the model transfers to scATAC-seq gene activity matrices with a handful of examples and no architectural change.

#Technical Details

Inputs are constructed by ranking genes globally: the top 3,000 highly variable genes from each of 570 curated single-cell studies were aggregated and ordered by recurrence across datasets, giving a fixed reference vocabulary. For a given cell, the expressed genes from the top k of that list are concatenated in ranked order to form the cell sentence — an approach related to Cell2Sentence's rank-based text encoding. Base models are Qwen2.5-7B-Instruct and QwQ-32B-Instruct, chosen after testing Qwen and LLaMA variants at several scales; Qwen2.5-7B-Instruct outperformed its 14B and 32B siblings at lower cost. Training and evaluation use Pancancer38k, 37,187 cells aggregated from published pan-cancer studies, with a held-out 3,800-cell test set. On a PDAC dataset, accuracy improved with cell-sentence length and plateaued at k = 1,000-1,500. On a class-balanced test set the 7B and 32B models scored 0.72 and 0.75, against 0.46 for DeepSeek-V3-671B and 0.52 for DeepSeek-R1-671B. On an out-of-distribution liver dataset containing unseen types such as hepatic stellate and γδ T cells, zero-shot accuracy was 0.33; 1-shot fine-tuning raised it to 0.73 and 10-shot to 0.93. Cross-modality transfer to scATAC-seq reached 0.53 at 1-shot and 0.65 at 10-shot without degrading scRNA-seq performance. Scalability was assessed on 181 public datasets spanning 48 cancer types from TISCH2.

#Applications

The intended use is routine annotation of new single-cell experiments, particularly in tumour atlases where subtype granularity matters and cluster-level labelling blurs distinctions. Because weights are released under Apache-2.0 and run locally on one GPU, the workflow avoids the cost, reproducibility, and data-privacy problems of routing patient-derived expression through a proprietary API — the main practical objection to GPT-based annotation. The reasoning traces double as review material: a curator can check whether a label rests on the markers they would have used. Released checkpoints include the base 7B and 32B models plus liver-adapted and scATAC-adapted 10-shot variants.

#Impact

CellReasoner is evidence that expert-level biological annotation requires neither a very large model nor a very large training corpus — 380 exemplars over a 7B open base beat 671B general reasoning models here. That reframes the cost structure of applying LLMs to omics interpretation. The reported limitations are dataset-dependent: zero-shot performance on genuinely unseen cell types is poor (0.33 on liver) and only recovers with few-shot examples, and accuracy varies widely even within a cancer type, from 0.95 to 0.48 across two multiple-myeloma datasets, reflecting inconsistent granularity in the reference labels as much as model failure. The work was published in the Journal of Genetics and Genomics in 2025. Code is on GitHub under Apache-2.0 and the HuggingFace repositories carry model cards; there is no separate data card, as the training corpora come from published, individually cited studies.