Cell Entity Annotation (CEA) Overview

• Cell Entity Annotation (CEA) is the process of assigning semantically meaningful labels to data units, such as table cells, cell clusters, or image pixels, drawn from predefined ontologies.
• CEA methods employ diverse strategies—from LLM-driven tool selection and probabilistic classifiers to multi-attribute image labeling—to address domain-specific challenges and enhance annotation accuracy.
• Evaluations utilize metrics like precision, recall, F1, BLEU scores, and image segmentation accuracy, demonstrating significant efficiency gains and improved performance across bioinformatics applications.

Cell Entity Annotation (CEA) is not a single, uniform task across the literature. In Semantic Table Annotation, it seeks a mapping from table cells to ontology entities, formally $f:E\to 2^{Ent}$ for the set of cell values $E$ and a target knowledge graph such as DBpedia (Geng et al., 18 Aug 2025). In single-cell transcriptomics, it can be viewed as assigning to each cell—or to each cell cluster—a discrete cell-type label drawn from a predefined ontology, or as producing a mapping $$\text{CEA}(x_i)\in\{1,\dots,K\}\cup\{\texttt{unknown}\}$$ when open-set recognition is required (Zeng et al., 2023, Mao et al., 7 Apr 2025). In microscopy, the same term has been used both for assigning a semantic class label to every pixel in a heterogeneous cell image and for assigning cell-type labels together with fine-grained morphological attributes to single-cell images (Naouar et al., 2023, Houmaidi et al., 30 Sep 2025). This terminological spread suggests that CEA is best understood as a family of entity-grounding problems whose common structure is the assignment of semantically meaningful labels to “cells,” but whose inputs, supervision, ontologies, and evaluation criteria differ substantially.

1. Semantic scope and core formalizations

In the table-annotation literature, CEA is defined over a table $T$ with $m$ rows and $n$ columns, where $E=\{e_{ij}\mid 1\le i\le m,\,1\le j\le n\}$ denotes the set of all cell values; the objective is to predict, for each cell $e_{ij}$, a possibly singleton set of ontology entities $E_{ij}=f(e_{ij})=\{e_1,e_2,\dots,e_b\}$ (Geng et al., 18 Aug 2025). The same paper places CEA inside the broader Semantic Table Annotation task, alongside Column Type Annotation.

In single-cell transcriptomics, one formulation uses a normalized expression matrix $X\in\mathbb{R}^{N\times G}$ for $E$0 cells and $E$1 genes, with Cell Entity Annotation defined as a mapping $E$2 from clusters to cell-type labels, or equivalently as a probabilistic classifier $E$3 with assignment $E$4 (Zeng et al., 2023). A more explicit open-set formulation defines $E$5, thereby making novelty detection part of the annotation problem (Mao et al., 7 Apr 2025). Cross-domain annotation between scRNA-seq and snRNA-seq has also been formalized as learning from a labeled source domain $E$6 and an unlabeled target domain $E$7, where the unknown target label set $E$8 induces a partial-domain-adaptation problem (Chen et al., 12 Nov 2025).

In microscopy, CEA has two distinct formal meanings. One assigns a semantic class label to every pixel in a microscopy image containing a heterogeneous mixture of cell types, distinguishing it from instance segmentation by requiring both “where are the cells” and “what type is each cell” (Naouar et al., 2023). Another defines CEA as the automated assignment of both cell-type labels and fine-grained morphological attributes to single-cell microscopy images, yielding a multi-attribute output rather than a single class label (Houmaidi et al., 30 Sep 2025).

2. CEA in semantic table annotation

For semantic tables, the central difficulties are semantic loss of cell values, strict ontology requirements, homonyms, spelling errors, and abbreviations (Geng et al., 18 Aug 2025). The LLM-agent approach in “An LLM Agent-Based Complex Semantic Table Annotation Approach” addresses these with a ReAct-based agent that iteratively reasons with an LLM and invokes external tools when needed. Five tools are described: a Data Preprocessing Tool, a Column Topic Detection Tool, a Knowledge Graph-Based Enhancement Tool, a Context-Supported CEA Selection Tool, and a Context-Supported CTA Selection Tool (Geng et al., 18 Aug 2025).

The table-CEA workflow is explicitly context-sensitive. If the column header is non-semantic but cell values are informative, the system invokes Column Topic Detection first and then proceeds to CEA; if cells are empty or meaningless but the header is clear, it skips CEA entirely and performs only CTA with richer context; otherwise, for a fully semantic column, it goes directly to KG lookup, CEA selection, and CTA ranking (Geng et al., 18 Aug 2025). The disambiguation step is phrased as selecting a candidate entity for a cell by using row context, the column name or inferred topic, and a ranked list of candidate entities returned by DBpedia lookup.

Efficiency is improved through distance-based reuse of existing annotations. The method defines a Levenshtein distance $E$9 recursively and reuses an existing annotation whenever

$$\text{CEA}(x_i)\in\{1,\dots,K\}\cup\{\texttt{unknown}\}$$0

Combined with deduplication, early stopping on high-confidence decisions, and caching of KG lookup calls, this reduces processing from 177,355 cells in a naïve pass to 60,341 distinct cells, yielding a $$\text{CEA}(x_i)\in\{1,\dots,K\}\cup\{\texttt{unknown}\}$$1 reduction in end-to-end annotation time and a $$\text{CEA}(x_i)\in\{1,\dots,K\}\cup\{\texttt{unknown}\}$$2 drop in total LLM tokens consumed (Geng et al., 18 Aug 2025).

Evaluation follows SemTab practice using

$$\text{CEA}(x_i)\in\{1,\dots,K\}\cup\{\texttt{unknown}\}$$3

On Tough Tables, the reported CEA performance is $$\text{CEA}(x_i)\in\{1,\dots,K\}\cup\{\texttt{unknown}\}$$4 and $$\text{CEA}(x_i)\in\{1,\dots,K\}\cup\{\texttt{unknown}\}$$5; on BiodivTab, it is $$\text{CEA}(x_i)\in\{1,\dots,K\}\cup\{\texttt{unknown}\}$$6 and $$\text{CEA}(x_i)\in\{1,\dots,K\}\cup\{\texttt{unknown}\}$$7 (Geng et al., 18 Aug 2025). Because the method is built around dynamic tool selection rather than a fixed matching rule, a plausible implication is that table CEA increasingly treats entity linking as a context-conditioned reasoning task rather than a pure lookup problem.

3. CEA in single-cell transcriptomics: supervised, integrative, and cross-domain settings

In single-cell RNA-seq, a standard pipeline begins with normalization, feature selection, dimensionality reduction, and clustering, after which marker genes $$\text{CEA}(x_i)\in\{1,\dots,K\}\cup\{\texttt{unknown}\}$$8 are obtained by differential expression and annotation is expressed as

$$\text{CEA}(x_i)\in\{1,\dots,K\}\cup\{\texttt{unknown}\}$$9

where $T$0 may include tissue, species, or developmental stage (Zeng et al., 2023). Prompt-based LLM annotation of clusters has been reported to use top markers and metadata, with benchmarking against existing tools such as scBERT, SingleR, and manual annotation (Zeng et al., 2023).

Cross-domain CEA between scRNA-seq and snRNA-seq introduces a distinct challenge: the target label set may be only a subset of the source label set. ScNucAdapt is described as the first method designed for cross-annotation between scRNA-seq and snRNA-seq datasets, and it addresses both distributional differences and cell-composition differences with partial domain adaptation (Chen et al., 12 Nov 2025). Its architecture has three modules: a shared encoder $T$1, a dynamic clustering and partial-match module, and a shared classifier $T$2. The encoder is a two-layer MLP,

$T$3

and the classifier is another two-layer MLP that maps $T$4 to $T$5 followed by softmax (Chen et al., 12 Nov 2025).

The training objective combines source-domain cross-entropy and a cluster-wise alignment term based on Cauchy–Schwarz divergence:

$T$6

The partial-match loss is defined over dynamically inferred target clusters $T$7 and matched source-label subsets $T$8,

$T$9

and dynamic split-and-merge rules, adopted from DeepDPM, determine the target-cluster count $m$0 (Chen et al., 12 Nov 2025). Only source classes with a matched target cluster contribute to $m$1, thereby ignoring nonshared labels and avoiding negative transfer. Reported accuracies include 90.24 on Bladder-immune (partial), 97.69 on Bladder-stromal (closed), 87.12 on Kidney unpaired (partial), 98.39 on Tumor-CLL (partial), 94.06 on Tumor-MBC (partial), and 99.78 on Mouse cortex (closed); removing either the CS-divergence loss or the dynamic clustering causes a substantial drop of 5–15 percentage points in accuracy across tasks (Chen et al., 12 Nov 2025).

A separate line of work addresses inconsistent label granularity across datasets. Motwani et al. formulate integrative cell type annotation with a finest-resolution set $m$2 and dataset-specific binning functions $m$3, where each observed label corresponds to a subset $m$4 of fine types (Motwani et al., 2021). Instead of discarding coarse labels, the model directly uses

$m$5

with a group-lasso penalty on gene coefficients and a ridge penalty on batch-specific parameters (Motwani et al., 2021). The resulting estimator is optimized by a blockwise proximal gradient descent algorithm. In simulation and real PBMC analyses, the integrative methods IBMR-int and IBMR-NG are reported to outperform subset and relabel baselines, often approaching an oracle estimator (Motwani et al., 2021). This suggests that, in transcriptomic CEA, label-resolution mismatch is not merely a nuisance variable but a modeling target in its own right.

4. Zero-shot, retrieval-augmented, and agentic CEA in single-cell analysis

Zero-shot single-cell CEA has been benchmarked as a prompt-based reasoning problem in which top marker genes are serialized into text and a model predicts the cell type directly or via zero-shot Chain-of-Thought prompting (Liu et al., 2024). SOAR evaluates 8 instruction-tuned LLMs across 11 datasets. On the scRNA-seq “Nature” benchmark, GPT-4o reports an average BLEU of 51.79 in zero-shot mode, while Mixtral-8×22B rises from 16.65 to 28.19 under zero-shot CoT; in the multiomics “scACT” benchmark, cross-modality translation through a VAE alignment module maps ATAC into pseudo-RNA before top-$m$6 marker extraction (Liu et al., 2024).

Agentic variants explicitly couple LLM planning with tool use. scAgent defines $m$7, where $m$8 is a planning module, $m$9 is an action space of scRNA models, modular Mixture-of-Experts LoRA adapters, embedding analysis, and incremental training procedures, and $n$0 stores reference datasets, vector embeddings in Milvus, and system history (Mao et al., 7 Apr 2025). Its MoE-LoRA update is

$n$1

and novel-cell discovery uses Euclidean-distance outlier detection and cosine similarity thresholds. On the CG benchmark, scAgent achieves macro $n$2, reported as $n$3 percentage points over the second-best scTab(10X), with 10× less data (Mao et al., 7 Apr 2025).

CellMaster frames zero-shot annotation as a collaborative loop involving a Hypothesis Generation Agent, a Marker Selection Agent, an Expression Analysis Module, and a Result Evaluation Agent, all operating on clustered AnnData inputs (Wang et al., 12 Feb 2026). It defines a cluster-type confidence score $n$4 by combining marker expression strength, specificity, and a dataset-wide $n$5-score stabilization term. Across 9 datasets spanning 8 tissues, CellMaster reports an average CL score of $n$6 in automatic mode versus a best baseline average of 0.531, and $n$7 with human-in-the-loop refinement, with a subtype gain of $n$8 and paired Wilcoxon signed-rank $n$9 after FDR correction (Wang et al., 12 Feb 2026).

Knowledge-graph retrieval is a second major direction. ReCellTy uses a graph-structured feature-marker database $E=\{e_{ij}\mid 1\le i\le m,\,1\le j\le n\}$0 with seven node types—Marker, FeatureFunction, CellName, CellType, TissueClass, CancerType, and GeneFamily—and five chained LLM-agent tasks: CellType Query, CellType Selection, Feature Query, Feature Selection, and CellType Annotation (Han et al., 24 Apr 2025). Reported gains include an average human-evaluation improvement of $E=\{e_{ij}\mid 1\le i\le m,\,1\le j\le n\}$1 over four baseline models, a $E=\{e_{ij}\mid 1\le i\le m,\,1\le j\le n\}$2 gain for DeepSeek-chat, and an overall semantic-evaluation improvement of $E=\{e_{ij}\mid 1\le i\le m,\,1\le j\le n\}$3, with GPT-4o-mini reaching up to $E=\{e_{ij}\mid 1\le i\le m,\,1\le j\le n\}$4 (Han et al., 24 Apr 2025).

GATHER introduces a more formal retrieval setting for what it calls zero-shot Cell Entity Annotation, where the query is a hyper-entity gene set $E=\{e_{ij}\mid 1\le i\le m,\,1\le j\le n\}$5 and the signal “emerges” from joint co-occurrence rather than any decisive single gene (Zhang et al., 7 May 2026). On a cell-centric biological knowledge graph with $E=\{e_{ij}\mid 1\le i\le m,\,1\le j\le n\}$6, $E=\{e_{ij}\mid 1\le i\le m,\,1\le j\le n\}$7, and approximately 14 relation types, GATHER performs global multi-source traversal, context-aware gene weighting, and topology-aware convergence scoring,

$E=\{e_{ij}\mid 1\le i\le m,\,1\le j\le n\}$8

With $E=\{e_{ij}\mid 1\le i\le m,\,1\le j\le n\}$9 and $e_{ij}$0, it reports exact-match accuracies of 27.45 on Immune and 59.64 on Lung, using only one LLM call per sample versus 2–61 calls for KG-RAG baselines, and improvements that are significant at $e_{ij}$1 against the best KG-RAG baseline on both datasets (Zhang et al., 7 May 2026).

Trustworthiness-oriented systems add explicit verification. CellTypeAgent uses an LLM to propose top candidate cell types and then verifies them with CellxGene-derived ranks based on scaled average expression, fraction of cells expressing each marker, and a tissue-agnostic global expression statistic (Chen et al., 13 May 2025). The final score is

$e_{ij}$2

The paper reports that CellTypeAgent consistently outperforms GPTCellType, CellxGene-only, and PanglaoDB, with representative average agreement scores of approximately 0.85, 0.78, 0.73, and 0.69, respectively (Chen et al., 13 May 2025). A common concern in this line of work is hallucination; the verification step is explicitly designed to use LLM outputs as hypotheses rather than final answers.

5. Imaging-based CEA: semantic segmentation and multi-attribute labeling

In heterogeneous microscopy images, CellMixer defines CEA as per-pixel semantic labeling rather than instance detection (Naouar et al., 2023). The method avoids pixel-level annotation by training on homogeneous cell populations with image-level labels and synthesizing heterogeneous mixtures through mixup-style augmentation. Foreground masks are obtained from a homogeneous image $e_{ij}$3 by computing Sobel gradients $e_{ij}$4, the gradient magnitude

$e_{ij}$5

followed by Gaussian smoothing, morphological erosion, adaptive thresholding, and assignment of the global class label to all foreground pixels (Naouar et al., 2023).

The mixed training sample combines two crops $e_{ij}$6 and $e_{ij}$7 with coefficient $e_{ij}$8, using $e_{ij}$9 in practice. CellMixer employs a DINOv2-pretrained Vision Transformer encoder and a Segmenter decoder head, with per-pixel logits over background and cell types and a Tversky loss to handle class imbalance (Naouar et al., 2023). On artificial mixtures, CellMixer improves over the baseline from 72.79/64.94 to 90.41/83.07 on Jurkat, from 88.20/80.44 to 96.68/89.87 on K562, and from 56.19/39.43 to 91.34/72.28 on PBMC, measured as mean pixel-accuracy and mean IoU; gains are also reported on real Jurkat+K562 and PBMC+Jurkat/K562 mixtures (Naouar et al., 2023).

AttriGen broadens the imaging definition by treating CEA as simultaneous cell-type and morphological-attribute annotation (Houmaidi et al., 30 Sep 2025). Its dual-model architecture combines a VGG16 CNN for 8-way cell-type classification on the PBC dataset and a Vision Transformer for 11-way multi-attribute classification on the WBCAtt dataset, with fusion by concatenation of the 8-dimensional probability vector and the 11-dimensional attribute vector into a 19-dimensional profile (Houmaidi et al., 30 Sep 2025). The type branch uses multi-class cross-entropy,

$E_{ij}=f(e_{ij})=\{e_1,e_2,\dots,e_b\}$0

while the attribute branch uses binary cross-entropy over the 11 attributes (Houmaidi et al., 30 Sep 2025).

The PBC dataset contains 17,092 images across eight classes, and WBCAtt contains 10,298 images with eleven morphological attributes (Houmaidi et al., 30 Sep 2025). Reported performance includes 98.83% accuracy for VGG16 on PBC and a new state-of-the-art GAA of 94.62% for Swin-S on WBCAtt, with highest per-attribute accuracies of 99.81% for Granularity, 99.61% for Granule Type, and 99.29% for Granule Color (Houmaidi et al., 30 Sep 2025). The paper further reports annotation at $E_{ij}=f(e_{ij})=\{e_1,e_2,\dots,e_b\}$1 ms/image, 2.26 minutes total for 6,784 unlabelled PBC cells, a 1.48% drop relative to a human expert accuracy of 96.10%, and an estimated cost reduction of $E_{ij}=f(e_{ij})=\{e_1,e_2,\dots,e_b\}$2 (Houmaidi et al., 30 Sep 2025). Here, CEA is tied not only to classification but also to interpretability, via Grad-CAM and attention-map analyses.

6. Evaluation regimes, recurrent limitations, and research directions

Because CEA spans distinct tasks, the evaluation regime depends on the object being annotated. Table CEA commonly uses precision, recall, and $E_{ij}=f(e_{ij})=\{e_1,e_2,\dots,e_b\}$3 (Geng et al., 18 Aug 2025). Single-cell transcriptomic studies variously report exact-match accuracy, ancestor-match accuracy in the Cell Ontology DAG, overall accuracy, weighted $E_{ij}=f(e_{ij})=\{e_1,e_2,\dots,e_b\}$4, macro $E_{ij}=f(e_{ij})=\{e_1,e_2,\dots,e_b\}$5, agreement score, and ontology-based CL score (Zhang et al., 7 May 2026, Mao et al., 7 Apr 2025, Chen et al., 13 May 2025, Wang et al., 12 Feb 2026). Imaging studies use mean pixel-accuracy, mean IoU, ordinary accuracy, and Global Average Accuracy (Naouar et al., 2023, Houmaidi et al., 30 Sep 2025). This suggests that cross-paper comparisons are only meaningful within a shared task definition and ontology regime.

The limitations reported in these literatures are also task-specific. For convergence-centric retrieval, missing or mis-grounded CellType nodes degrade performance, a fixed horizon $E_{ij}=f(e_{ij})=\{e_1,e_2,\dots,e_b\}$6 may miss longer but informative indirect paths, and retrieval parameters such as $E_{ij}=f(e_{ij})=\{e_1,e_2,\dots,e_b\}$7 and $E_{ij}=f(e_{ij})=\{e_1,e_2,\dots,e_b\}$8 require tuning because there is no end-to-end training (Zhang et al., 7 May 2026). For LLM-based single-cell annotation, hallucination, prompt sensitivity, latency, and cost remain recurrent concerns; CellTypeAgent notes that literature-search augmentation often hurts performance and increases prompt length, while CellMaster and scAgent note reliance on large LLMs and careful prompt engineering (Chen et al., 13 May 2025, Wang et al., 12 Feb 2026, Mao et al., 7 Apr 2025). In transcriptomic open-set settings, novel or ultra-rare cell types absent from verification databases cannot be validated directly (Chen et al., 13 May 2025). In imaging, synthetic mixing assumes linear blending and does not model occlusion or depth ordering in tightly packed clusters, and very dense clusters remain challenging (Naouar et al., 2023).

Several forward directions recur across the papers. GATHER proposes dynamic horizon and relation-type weighting, joint training of retrieval scoring parameters, integration with supervised classifiers such as scGPT, and application to other hyper-entity reasoning domains (Zhang et al., 7 May 2026). scAgent proposes extending the agent to multi-omic, spatial, and perturbation data via additional LoRA plugins (Mao et al., 7 Apr 2025). SOAR recommends incorporating external cell-marker databases via RAG and extending multiomics evaluation beyond ATAC (Liu et al., 2024). CellTypeAgent suggests dynamically weighting LLM and database evidence, expanding verification to multiple databases, and improving mixture handling through structured multi-label prompts and scoring (Chen et al., 13 May 2025). Across these works, a consistent pattern is that CEA increasingly combines structured retrieval, ontology grounding, and adaptive reasoning, but the precise meaning of “entity” remains domain-dependent.