CellMaster: Collaborative Cell Type Annotation in Single-Cell Analysis

Zhen Wang; Yiming Gao; Jieyuan Liu; Enze Ma; Jefferson Chen; Mark Antkowiak; Mengzhou Hu; JungHo Kong; Dexter Pratt; Zhiting Hu; Wei Wang; Trey Ideker; Eric P. Xing

Source-linked AI summary

CellMaster: Collaborative Cell Type Annotation in Single-Cell Analysis

Zhen Wang, Yiming Gao, Jieyuan Liu, Enze Ma, Jefferson Chen, Mark Antkowiak, Mengzhou Hu, JungHo Kong, Dexter Pratt, Zhiting Hu, Wei Wang, Trey Ideker, Eric P. Xing

arXiv:2602.13346v1Feb 12, 2026q-bio.GNcs.AIcs.CV

TL;DR

Cell-type annotation in scRNA-seq remains difficult because marker usage is context-dependent and rare or novel states often lack consensus markers. CellMaster uses an LLM-driven, zero-shot expert-like reasoning process and outperformed the best dataset-specific baselines by 0.071 on average across 9 datasets.

Problem
ScRNA-seq cell-type annotation remains challenging because marker usage is context-dependent and rare, transitional, or novel states lack consensus markers.
Method
CellMaster is an LLM-driven agent that interprets cluster-level differential-expression markers and dataset context to generate zero-shot, explainable annotations without pretraining or fixed marker tables.
Results
Across 9 datasets, CellMaster outperformed each dataset’s best baseline by 0.071 on average, achieving 0.602 ± 0.058 performance.
Takeaways & Limitations
Human feedback improved average performance by 0.115, highlighting the value of expert intuition in guiding collaborative annotation.
Takeaways & Limitations
Evaluation based on Cell Ontology mapping may bias assessment against genuinely novel cell states absent from existing hierarchies.

Abstract

from arXiv · show

Single-cell RNA-seq (scRNA-seq) enables atlas-scale profiling of complex tissues, revealing rare lineages and transient states. Yet, assigning biologically valid cell identities remains a bottleneck because markers are tissue- and state-dependent, and novel states lack references. We present CellMaster, an AI agent that mimics expert practice for zero-shot cell-type annotation. Unlike existing automated tools, CellMaster leverages LLM-encoded knowledge (e.g., GPT-4o) to perform on-the-fly annotation with interpretable rationales, without pre-training or fixed marker databases. Across 9 datasets spanning 8 tissues, CellMaster improved accuracy by 7.1% over best-performing baselines (including CellTypist and scTab) in automatic mode. With human-in-the-loop refinement, this advantage increased to 18.6%, with a 22.1% gain on subtype populations. The system demonstrates particular strength in rare and novel cell states where baselines often fail. Source code and the web application are available at \href{https://github.com/AnonymousGym/CellMaster}{https://github.com/AnonymousGym/CellMaster}.

1 Introduction

Cell-type annotation in scRNA-seq is difficult because marker usage depends on tissue and cellular state, while rare, transitional, and novel populations often lack consensus markers. CellMaster addresses this gap with a zero-shot, explainable LLM agent that reasons over markers in context and iteratively refines uncertain labels without pretraining or fixed marker tables.

Motivation: Marker usage is highly context-dependent across fine-grained lineages and transient cellular states in complex tissues.Examples include monocyte and T-cell subtypes, hepatoblasts, intermediate mature neutrophils, and pre-B cells.
Motivation: Expert annotation can produce high-quality labels but is time-consuming, scales poorly, and is especially challenging for rare, transitional, or novel states.Biological interpretation is required to connect molecular signatures with true cellular identities.
Related work: Automated methods based on marker matching, reference transfer, or supervised classification depend on reference coverage and nomenclature alignment, making them vulnerable to domain shifts and novel states.Such methods may abstain or misassign when confronted with rare or truly novel populations.
Research gap: A task-grounded system combining in-situ marker reasoning, biologically interpretable explanations, and zero-shot operation without task-specific pretraining or fixed marker tables had not been established.LLMs and human feedback had shown promise, but the stated combination remained a gap for scRNA-seq annotation.
Contribution: CellMaster mirrors an expert reasoning loop by interpreting cluster-level DE markers with dataset context, generating natural-language rationales, and iteratively testing hypotheses to resolve near-neighbor confusions.The system performs biology-aware refinement and escalates uncertain clusters for further handling.

2 Related Works

Related work spans gene-token foundation models and language-based single-cell analysis, while newer biomedical agents combine LLM reasoning with tool use. CellMaster differs by enabling iterative human intervention rather than delivering only fixed, closed-pipeline outputs.

2 Related Works: Single-cell foundation models, including scBERT, scGPT, and Geneformer, learn latent gene representations for tasks such as imputation and perturbation prediction.These models represent one major direction in applying LLM-related methods to single-cell analysis.
2 Related Works: Embedding-based models hinder interpretability, while language interfaces often remain thin wrappers around predefined pipelines with limited reasoning or adaptability.Recent work therefore shifts toward autonomous agents combining LLM reasoning with tool use.
2 Related Works: CASSIA, scRAG, and scPilot emphasize automated annotation or cross-tissue generalization but typically provide finalized results through closed pipelines with limited user intervention.CASSIA uses an Annotator–Validator–Scorer assembly line, whereas scRAG augments reasoning with retrieved structured knowledge such as cell–tissue relationships.
2 Related Works: CellMaster prioritizes iterative human-in-the-loop interaction, allowing users to adjust annotation granularity, critique intermediate reasoning, and refine visual interpretations.Its open-loop design positions the LLM as a responsive research partner aligned with expert biological intuition and hypothesis-driven exploration.

3 Methodology

CellMaster iteratively refines cluster annotations through hypothesis generation, adaptive marker selection, expression analysis, and multilevel evaluation. Its ontology-based benchmark enables biologically coherent comparisons, while the multi-agent design supports iterative zoom-in refinement.

Annotation pipeline: CellMaster iteratively refines cluster annotations through four specialized stages, with detailed prompts provided in Supplement G.The pipeline represents cells as x_i ∈ R^M and maps clusters to cell types or undefined annotations at each iteration.
Annotation pipeline: Hypothesis refinement uses Wilcoxon rank-sum differential expression analysis to select the top N genes by average log-fold change for each cluster.These genes provide the foundation for refining cell-type hypotheses across iterations.
Annotation pipeline: Adaptive marker proposal maintains successful and failed marker genes across iterations, while dotplots verify whether proposed markers discriminate cell types and reveal distinct expression patterns.The system uses marker memory, confidence stabilization, cluster relationship analysis, and contamination detection heuristics.
Evaluation framework: Ontology-based evaluation standardizes marker names, maps cell types to ontologies, constructs hierarchical annotations, and quantifies agreement with reference standards.The framework is designed to support biologically coherent comparisons between predicted annotations and established references.
Benchmark design: Large biological language models were excluded because tissue-composition bias can limit generalizability and fine-tuning requires substantial GPU resources.The cited example reports 71% of scGPT’s training cells from blood and brain tissues, while fine-tuning approximately 10,000 cells requires at least 8GB of GPU memory post-optimization.
Method comparison: CellMaster uses a multi-agent LLM structure to evaluate marker expression within clusters and supports iterative zoom-in refinement beyond initial clustering.Unlike the baselines, annotations are based primarily on initial clustering but can be refined for enhanced granularity.

4 Results

CellMaster combines automated and human-in-the-loop annotation with iterative, context-aware reasoning and interpretable outputs. Across diverse datasets, it outperformed baselines, improved with minimal human feedback, and generated biologically plausible subtype insights.

Framework: CellMaster supports automated and human-in-the-loop annotation through a multi-agent LLM architecture built on Scanpy.The two modes combine automated efficiency with researcher-guided precision.
Framework: Each iteration generates hypotheses, proposes markers, synthesizes evidence, and returns labels or abstentions with rationales, visualizations, confidence, and next steps.The workflow prioritizes biological plausibility and interpretability within experimental context.
Automatic performance: 0.602 ± 0.058 average performance across 9 datasets exceeded each dataset’s best baseline by 0.071 on average.The largest reported gains were in Retina (0.705 vs 0.300 - 0.632) and Liver (0.55 vs 0.304 - 0.429).
Automatic performance: CellMaster outperformed GPTCelltype on 8/9 datasets while showing modest variability with an average s.d. ≈0.058.It also consistently outperformed Biomni in accuracy and stability on representative datasets.
Fine-grained analysis: CellMaster consistently exceeded GPTCelltype and CellTypist across most broad cell types, with substantial advantages for T cells and neural populations.It matched baseline performance for B cells and NK cells and outperformed baselines across varied dataset and cluster-size conditions.
Human-in-the-loop refinement: 0.115 ± 0.085 accuracy improvement resulted from human-in-the-loop feedback under a 3 interaction budget compared with automatic CellMaster.Brief expert guidance redirected analysis toward ambiguous regions and reduced false positives.

5 Discussions

CellMaster’s discussion emphasizes adaptive LLM reasoning and interpretable rationales as advantages over static annotation models. Human feedback improves performance, while ablations reveal performance plateaus and drops without feedback, underscoring continued reliance on expert oversight.

Adaptive reasoning: CellMaster dynamically adapts to novel annotation contexts without retraining and provides interpretable rationales, unlike static models such as CellTypist and CellMarker 2.0.The discussion cites traced reasoning, including Neutrophil subtype differentiation in a Liver case study.
Human–AI collaboration: 0.115 average performance gain with human feedback highlights the critical role of expert intuition, supported by a UI embedding biologist workflows for collaboration.The interface is contrasted with ‘black-box‘ tools.
Limitations: After 3–5 iterations, ablations show performance plateaus and drops in no-feedback mode, underscoring AI’s reliance on human oversight for nuanced decisions.The discussion links oversight to preventing model overconfidence, a challenge also observed in other LLM applications.

6 Limitations

CellMaster remains limited by LLM stochasticity, commercial API dependence, and potential evaluation bias against genuinely novel cell states. These limitations affect reproducibility, deployment in regulated settings, and assessment validity.

Reproducibility: 0.058 mean s.d. reflects higher run-to-run variability than GPTCelltype’s 0.053, despite stabilization mechanisms.The probabilistic nature of LLMs introduces stochasticity.
Deployment: Dependence on commercial model APIs raises cost and privacy concerns for regulated settings.This dependence may constrain deployment where cost control or data privacy is required.
Evaluation: Cell Ontology mapping may bias evaluation against genuinely novel cell states absent from the mapping.Novel states without corresponding ontology entries may be assessed less favorably.

7 Conclusion

CellMaster combines LLM adaptive reasoning with biologist expertise through an iterative interface for versatile single-cell annotation. Its performance across tissues and subtypes, handling of edge cases, and emphasis on interpretability and collaboration support AI adoption in life sciences.

CellMaster synergizes LLMs’ adaptive reasoning with biologist expertise through an intuitive, iterative interface.
CellMaster outperforms existing tools across major tissues and subtypes while robustly handling edge cases.
The framework is positioned as a versatile solution for diverse single-cell studies.
Interpretability and human collaboration address critical barriers to AI adoption in life sciences.

8 Ethics and Data Use

CellMaster uses only publicly available, anonymized datasets from prior publications and generates no new human-subjects data. Accordingly, the work is a secondary analysis of de-identified data exempt from IRB oversight.

8 Ethics and Data Use: The study exclusively analyzes publicly available, anonymized datasets derived from prior publications.No new human subjects data were generated, collected, or processed.
8 Ethics and Data Use: The work is a secondary analysis of de-identified data and is exempt from Institutional Review Board oversight.

9 Data and Code Availability · A Additional Related Works

CellMaster’s benchmarking resources include publicly sourced processed datasets, cell-type labels, and openly released code. Related work highlights the growing complexity of single-cell analysis and the scalability and reference-bias limitations of existing annotation approaches.

9 Data and Code Availability: CellMaster releases publicly sourced processed scRNA-seq datasets in H5AD format, DEG data in CSV format, and accompanying cell-type labels.These collections are the processed datasets used for CellMaster benchmarking.
9 Data and Code Availability: The complete CellMaster source code, preprocessing pipeline, evaluation metrics, and benchmark drivers are available under the MIT license.The repository is available at https://github.com/AnonymousGym/CellMaster.
A Additional Related Works: Single-cell RNA sequencing has grown exponentially through large consortia such as the Human Cell Atlas, reaching datasets with millions of cells.This scale has driven new analytical demands in single-cell research.
A Additional Related Works: Analysis has shifted from coarse clustering toward resolving fine-grained and continuous cellular heterogeneity in increasingly large datasets.The related-work discussion frames this shift as a consequence of atlas-scale scRNA-seq.
A Additional Related Works: Cell identity is increasingly understood as a dynamic spectrum shaped by developmental trajectories, microenvironments, and disease states.The passage notes that computationally capturing this subtle structure remains difficult.
A Additional Related Works: Manual marker-based annotation is interpretable and accurate but unscalable, whereas reference-based methods such as SingleR and CellTypist are efficient yet constrained by reference bias.Reference-based approaches can only assign labels present in their training atlas, limiting novel or context-specific state annotation.

B Dataset Information Details … D.6 Summary

CellMaster was evaluated across nine scRNA-seq datasets spanning eight tissues and used adaptive heuristics plus ontology-based evaluation to improve annotation robustness and biological relevance. Its evaluation combined standardized naming, dynamic ontology mapping, hierarchical scoring, and contextual validation.

B Dataset Information Details: Nine diverse scRNA-seq datasets spanning eight tissues were used to assess CellMaster’s cross-tissue generalization and performance.The benchmark included human and mouse samples from liver, peripheral blood, lymphoma, bone marrow, brain, large intestine, and limb muscle.
C.1 Adaptive Heuristics: Marker Memory records successful and unsuccessful marker genes across iterations, with success defined by a combined expression score meeting success_threshold = 0.7.The combined score incorporates expression level and the fraction of expressing cells in a cluster.
C.1 Adaptive Heuristics: Confidence Stabilization preserves annotations when normalized-expression z-scores satisfy Z ≥2.0 and flags other clusters for refinement.The criterion is designed to support stable annotations across iterative processing.
C.1 Adaptive Heuristics: Cluster Relationship Analysis identifies similar expression profiles using differential-expression and fraction thresholds to guide targeted sub-clustering of ambiguous clusters.This analysis highlights cluster pairs with limited distinguishing genes and supports refined cell-type assignments.
C Supplementary Methods: Contamination Rules review clusters exceeding a doublet-ratio threshold or showing elevated mitochondrial fractions to protect annotation reliability and dataset integrity.The rules address potential doublets and mitochondrial contamination.
D.1 Cleaning and Standardizing Cell Type Names: Cell-type names were cleaned and standardized through singularization, removal of redundant whitespace and punctuation, retention of biologically relevant symbols, and dictionary or language-model mapping.The process harmonized names for consistent downstream analysis, including uncommon or ambiguous terms.
D.2 Mapping to Cell Ontology Identifiers: Standardized names were queried against the Cell Ontology to retrieve CLIDs and higher-level categories, while unmapped names were retained for review.This mapping aligned predicted and reference annotations with structured biological knowledge.

E Statistical Analysis … I Iterative Annotation Performance Dynamics

CellMaster combines statistically controlled evaluation with an iterative, multi-agent annotation workflow that supports user-defined granularity and human feedback. Performance improves through early iterations but declines with prolonged unattended cycling, motivating controlled refinement.

E Statistical Analysis: Differential expression used two-sided Wilcoxon rank-sum tests with Benjamini–Hochberg correction; significance required adjusted p value <0.05 and |log2 fold change| > 1.Analyses used Python 3.11 with Scanpy 1.10, SciPy 1.13, NumPy 1.26, and statsmodels 0.14.
E Statistical Analysis: Agreement scores used GPTCelltype’s ontology metric, averaged across clusters, with paired Wilcoxon signed-rank tests across nine datasets and 5% FDR control.Scores were 1.0 for exact matches, 0.5 for parent/child matches, and 0 for no matches; tests were two-sided unless noted.
F Comprehensive Feature Comparison with Existing Paradigms: CellMaster combines large language model availability with the user-defined annotation granularity typically reserved for manual curation.Users can guide resolution through natural-language distinctions such as “Immature Neutrophil” versus “Neutrophil,” unlike standard automated tools and rigid GPT scripts.
G Detailed workflow of CellMaster agents: The workflow generates dataset-informed hypotheses from top cluster markers, proposes markers for cell types of interest, and evaluates predictions using expression patterns and dotplots.Hypothesis generation can incorporate dataset context and previous iterations, while marker selection and evaluation support iterative refinement.
G Detailed workflow of CellMaster agents: Expression analysis identifies markers using mean-expression and detection-fraction thresholds, compares clusters with log-fold changes, and finds similar pairs using relaxed distinguishing-marker criteria.These functions return cluster-level marker lists and similar-cluster pairs based on configurable thresholds and maximum differing-gene counts.
H Details of UI Inputs and Outputs: Example: The interface exposes staged controls, hypothesis and marker outputs, expression and UMAP visualizations, cluster selection, and chat-based human feedback for subsequent iterations.Users can provide guidance such as “use Nkg7 for NK cell” and either submit feedback or proceed to the next iteration.
I Iterative Annotation Performance Dynamics: Accuracy rose monotonically from iteration 1 through iteration 4 on Liver, then declined at iteration 5 as iterative refinement continued beyond a productive point.The improvement reflected annotation of blank clusters, correction of early mislabels, and incorporation of new marker evidence.
I Iterative Annotation Performance Dynamics: Peak accuracy occurred between iterations 3 and 5 across PBMC, BCL, and Liver, while prolonged unattended cycling caused a mild but consistent decline.The downturn was attributed to over-iteration: self-correction without fresh external cues can amplify spurious signals among remaining ambiguous cells.

J Biological insight example of B-Cell subgrouping

CellMaster used iterative exploration to subdivide developmental liver B cells and frame their annotation around canonical developmental stages. Marker-enriched expression patterns supported interpretable subtype assignments and recapitulated the established B-cell trajectory without lineage-specific retraining or marker databases.

Exploration: CellMaster’s Evaluation module flagged developmental B-cell subtypes for further exploration after initial annotation iterations.The suggested context included developmental stages such as pre-B and pro-B cells.
Subclustering: User-guided analysis generated 6 B-cell sub-clusters at resolution 0.3 from clusters 0, 7, and 20.The Hypothesis module proposed pro-B, large pre-B, small pre-B, and naive B cells as canonical neonatal-liver stages.
Marker evidence: Ebf1 and Vpreb1 marked clusters 0 and 5, Cd79a/Cd79b/Igll1 marked cluster 1, Rag1 marked cluster 2, and Ms4a1 marked clusters 3–4.These dotplot patterns provided the expression basis for downstream subtype annotation.
Annotation: Clusters 0 and 5 were assigned as Pro-B Cells, while cluster 1 was assigned as Large Pre-B Cells using marker-based rationales.The reported rationales were Cd19, Ebf1, Il7r, and Vpreb1 for Pro-B Cells, and Cd79a, Cd79b, and Igll1 for Large Pre-B Cells.
Biological insight: The analysis recapitulated the established B-cell developmental trajectory in fetal and neonatal liver without lineage-specific retraining or marker databases.This result demonstrates generalization of CellMaster’s reasoning framework across immune lineages.