CellMaster: Collaborative Cell Type Annotation in Single-Cell Analysis

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}.

💡 Research Summary

The manuscript introduces CellMaster, an AI‑driven framework that tackles one of the most persistent bottlenecks in single‑cell RNA‑sequencing (scRNA‑seq) analysis: reliable cell‑type annotation. Traditional automated annotators rely on static marker gene lists or pre‑trained classification models, which become brittle when marker expression varies across tissues, developmental stages, or disease contexts, and they fail outright for novel or rare cell states lacking reference profiles. CellMaster circumvents these limitations by leveraging the encyclopedic knowledge embedded in large language models (LLMs), specifically GPT‑4o, to perform zero‑shot annotation without any prior training on the target dataset or a fixed marker database.

The workflow consists of four main stages. First, raw count matrices are normalized, dimensionally reduced (PCA/UMAP), and clustered using community‑detection algorithms such as Leiden. Second, for each cluster the system extracts a concise statistical summary: top differentially expressed genes, log‑fold changes, and adjusted p‑values. Third, this summary is inserted into a carefully crafted prompt template that asks the LLM to “suggest plausible cell types for this cluster and provide a rationale, citing key marker genes, tissue context, and functional attributes.” The LLM returns a ranked list of candidate labels together with an interpretable justification for each. Fourth, the results are stored in a database and displayed in an interactive web interface where users can accept, reject, or edit the suggestions. Any user feedback is fed back into the prompt generation step, enabling a human‑in‑the‑loop (HITL) refinement loop that iteratively improves annotation quality.

To mitigate the high computational cost of repeated LLM calls, CellMaster incorporates a caching layer that reuses prior responses for identical cluster signatures and a batch‑processing mode that aggregates multiple clusters into a single API request. Additionally, the authors provide an extensible library of domain‑specific prompt templates (e.g., immunology, neurobiology) and an automated prompt‑tuning routine that adjusts temperature, max‑tokens, and few‑shot examples based on user‑provided validation data.

The authors benchmarked CellMaster on nine publicly available scRNA‑seq datasets spanning eight distinct tissues (brain, liver, lung, heart, immune system, tumor, etc.). Baseline methods included CellTypist, scTab, SingleR, scPred, and a naïve marker‑gene lookup. In pure zero‑shot mode, CellMaster achieved an average overall accuracy improvement of 7.1 % relative to the best baseline, with particularly striking gains on rare populations (e.g., hepatic macrophage subtypes, embryonic neural progenitors) where baseline accuracies fell below 30 % but CellMaster exceeded 60 %. When expert users engaged in the HITL loop, overall accuracy rose to a 18.6 % improvement, and subtype discrimination (e.g., T‑cell exhaustion versus activation states) improved by 22.1 %. These results demonstrate that the LLM’s broad biomedical knowledge can fill gaps left by incomplete reference atlases, especially for novel or transient cell states.

The paper also discusses limitations. LLMs are trained on static corpora; consequently, very recent discoveries or highly specialized experimental conditions may be misinterpreted, leading to erroneous label suggestions. Moreover, API latency and token‑based pricing can become prohibitive for very large projects. The authors propose future directions such as fine‑tuning smaller, open‑source LLMs on curated cell‑type literature, integrating ontology‑aware post‑processing to enforce hierarchical consistency, and developing community‑driven prompt repositories to share best practices.

In conclusion, CellMaster represents a paradigm shift from static marker‑based annotation toward a dynamic, knowledge‑driven approach that couples the reasoning capabilities of LLMs with expert oversight. By eliminating the need for pre‑trained classifiers and enabling real‑time, interpretable rationales, it promises to accelerate atlas construction, facilitate discovery of disease‑specific cell states, and support precision‑medicine pipelines. The source code, Docker images, and a hosted web application are publicly released on GitHub (https://github.com/AnonymousGym/CellMaster), inviting immediate adoption and further community development.