PaperlayerView exact revision on arXiv

Source-linked AI summary

LLM4Cell: A Survey of Large Language and Agentic Models for Single-Cell Biology

Sajib Acharjee Dip, Adrika Zafor, Bikash Kumar Paul, Uddip Acharjee Shuvo, Muhit Islam Emon, Xuan Wang, Liqing Zhang

arXiv:2510.07793v3cs.CLcs.AI

TL;DR

Single-cell LLM research remains fragmented across modalities, architectures, and evaluation standards. LLM4Cell synthesizes models and public datasets in a shared taxonomy and evaluation rubric, revealing a shift toward language-driven and increasingly autonomous systems for single-cell analysis.

  • Problem

    Single-cell models vary in modality, supervision, and evaluation standards, making cross-model comparison and reproducibility difficult.

    Source: §1 · p. 1

  • Method

    LLM4Cell surveys 58 models and over 40 public datasets, organizing methods by family and task and evaluating them with a ten-dimension rubric.

    Source: §1 · p. 2, §4 · p. 3

  • Results

    The taxonomy and rubric reveal a transition from purely statistical modeling toward language-driven, interpretable, and increasingly autonomous systems connecting molecular data with biological knowledge.

    Source: §8 · p. 9

  • Takeaways & Limitations

    The synthesis is intended as a reproducible reference for benchmarking, model selection, and designing next-generation cellular foundation and reasoning models.

    Source: §8 · p. 9

    Reported metrics vary across publications, the rubric provides qualitative rather than standardized numerical rankings, and some clinical or proprietary spatial datasets were inaccessible.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Large language models (LLMs) and emerging agentic frameworks are beginning to transform single-cell biology by enabling natural-language reasoning, generative annotation, and multimodal data integration. However, progress remains fragmented across data modalities, architectures, and evaluation standards. LLM4Cell presents the first unified survey of 58 foundation and agentic models developed for single-cell research, spanning RNA, ATAC, multi-omic, and spatial modalities. We categorize these methods into five families-foundation, text-bridge, spatial, multimodal, epigenomic, and agentic-and map them to eight key analytical tasks including annotation, trajectory and perturbation modeling, and drug-response prediction. Drawing on over 40 public datasets, we analyze benchmark suitability, data diversity, and ethical or scalability constraints, and evaluate models across 10 domain dimensions covering biological grounding, multi-omics alignment, fairness, privacy, and explainability. By linking datasets, models, and evaluation domains, LLM4Cell provides the first integrated view of language-driven single-cell intelligence and outlines open challenges in interpretability, standardization, and trustworthy model development.

1 Introduction

Single-cell LLM research spans multiple modalities and evaluation standards, leaving models difficult to compare and reproducibility challenging. LLM4Cell surveys 58 methods, organizes them by family and task, and evaluates them alongside public datasets using a ten-dimension rubric.

1 Introduction

  • Fragmented modalities, supervision types, and evaluation standards make cross-model comparison and reproducibility difficult.The field spans scRNA-seq, ATAC, multi-omic, and spatial measurements, while agentic reasoning lacks standardized measures of correctness or biological grounding.Source: §1 · p. 1
  • LLM4Cell organizes 58 methods into five methodological families and eight analytical tasks, including annotation, trajectory inference, perturbation, and drug-response modeling.The families cover foundation, text-bridge, spatial/multimodal, epigenomic, and agentic approaches.Source: §1 · p. 1
  • The survey compiles over 40 public datasets across RNA, ATAC, multi-omic, spatial, perturbation, and plant domains.It evaluates models with a ten-dimension rubric covering grounding, fairness, scalability, and interpretability.Source: §1 · p. 2
  • LLM4Cell combines a dataset registry, model taxonomy, domain rubric, and discussion of challenges in cross-modal alignment, reasoning, and trustworthy AI.Its rubric also captures generalization and ethics.Source: §1 · p. 2

2 Related Work

Prior surveys and benchmarks cover architectures, prompting, and selected applications, but offer limited synthesis of datasets, evaluation standards, and agentic methods. LLM4Cell links models, datasets, and evaluation domains across modalities and tasks.

2 Related Work

  • Earlier reviews give little attention to datasets for spatial, perturbation, and multimodal modeling, while standardized benchmarks, domain-aware metrics, and reproducible splits are seldom addressed.Plant single-cell datasets and agentic reasoning as a coherent methodological class are also largely absent from prior reviews.Source: §2 · p. 2
  • LLM4Cell links over 40 public datasets and 58 models across modalities and tasks, assessing biological grounding, fairness, and scalability.The authors describe this data-centric synthesis as a reproducible foundation for cross-modal benchmarking and trustworthy, autonomous single-cell intelligence.Source: §2 · p. 2

3 Datasets

The survey compiles public datasets supporting pretraining, evaluation, and cross-modal reasoning across single-cell modalities and tasks. It identifies imbalances and access or metadata constraints that limit reproducible benchmarking.

3 Datasets

  • The survey compiles more than forty public resources across RNA, ATAC, multi-omic, spatial, perturbation/drug-response, and plant domains.These datasets support pretraining, evaluation, and cross-modal reasoning for tasks including annotation, integration, trajectory inference, perturbation modeling, and spatial mapping.Source: §3 · p. 2
  • RNA atlases dominate in scale, while chromatin-accessibility datasets remain sparse and heterogeneous and non-human and plant resources remain scarce.Multi-omic resources link RNA, ATAC, and protein modalities; spatial datasets connect molecular profiles with tissue architecture.Source: §3 · p. 2
  • Fragmented benchmarks and inconsistent metadata hinder reproducibility, while privacy constraints limit access to clinical spatial datasets.Paired and trimodal resources are identified as testbeds for next-generation multimodal and agentic LLMs.Source: §3 · p. 2
  • Appendix Tables 1–6 list representative datasets with tasks, scale, and source links for reproducible benchmarking.The passage identifies the tables but does not provide their individual dataset values.Source: §3 · p. 3

4 Model Taxonomy

LLM4Cell classifies single-cell models by modality, grounding, agentic capability, task, and domain quality. Its families range from foundation models to systems for spatial, epigenomic, and autonomous reasoning.

4.1 Foundation Models

  • Foundation models learn transferable cell and gene embeddings from large scRNA-seq datasets, typically using masked-gene or rank-based reconstruction.scGPT, Geneformer, and scFoundation (Hao et al., 2024) pretrain on multi-tissue atlases exceeding one million cells; these models support transfer for annotation and integration but have limited ontology grounding and explainability.Source: §4.1 · p. 3

4.2 Text-Bridge LLMs

  • Text-Bridge models align molecular embeddings with biomedical language or translate expression profiles into natural-language summaries.They use contrastive or prompt-based alignment, enhancing interpretability and enabling zero-shot annotation, but depend on curated corpora and remain non-agentic.Source: §4.2 · p. 3

4.3 Spatial and Multimodal Models

  • Spatial and multimodal models fuse gene expression with spatial coordinates, histology, or other omics to represent tissue architecture.They use modality-specific encoders and cross-attention, but face heterogeneous resolutions, limited open benchmarks, and high computational cost.Source: §4.3 · p. 3

4.4 Epigenomic Models

  • Epigenomic models learn chromatin-accessibility and regulatory patterns, with cross-omic variants jointly embedding RNA and ATAC to infer gene-regulatory networks.Sparse data and the lack of unified benchmarks across regulatory modalities constrain this family.Source: §4.4 · p. 3

4 Model Taxonomy

  • The taxonomy places 58 methods into five families and tracks modality, grounding, agentic capability, primary task, and ten domain-quality scores.Foundation models dominate scope and adoption, while spatial, epigenomic, and agentic approaches represent a shift toward contextual, reasoning-driven systems.Source: §4 · p. 3, Figure 1 · p. 4

4.5 Agentic Frameworks

  • Agentic systems combine domain-specific encoders with LLM controllers that plan tasks, query ontologies, and use tools or APIs for multi-step analysis.They support dialogue-based annotation and reasoning, but lack standardized reasoning-fidelity benchmarks and depend on underlying LLM reliability.Source: §4.5 · p. 4

5 Task-specific Applications

Across 58 methods, the section maps eight recurring tasks and shows how model families differ in operational scope, from common annotation and integration work to specialized spatial, regulatory, and agentic capabilities.

5.1 Annotation and Ontology Mapping

  • Annotation is the most common task, covering cell-type labeling, ontology alignment, and cross-dataset harmonization.Evaluation commonly uses label accuracy or ARI; rare-cell detection, cross-species consistency, and reasoning benchmarks remain limitations.Source: §5.1 · p. 5

5.2 Trajectory and Perturbation Modeling

  • Trajectory and perturbation models capture dynamic responses, but prediction degrades for combinatorial or long-range effects.Methods use denoising or contrastive objectives, with expression-profile correlation and target-gene recovery among the metrics.Source: §5.2 · p. 5

5.3 Multi-omic Integration

  • Multi-omic models align RNA, ATAC, and protein features, improving cell-type alignment and batch correction over unimodal models.Sparse modality overlap, data-scale imbalance, limited interpretability, and a lack of standard benchmarks constrain the approach.Source: §5.3 · p. 5

5.4 Spatial Mapping and Deconvolution

  • Spatial methods combine coordinates, histology, and gene-expression features to map tissue structure, but heterogeneous resolutions and limited benchmarks constrain comparisons.The accompanying distribution figure says spatial transcriptomics models most emphasize biological grounding, scalability, and annotation tasks.Source: Figure 2 · p. 5, §5.4 · p. 6

5.5 Regulatory Network and Pathway Inference

  • Regulatory and pathway models learn chromatin or gene dependencies, yet sparse training data and missing unified ground-truth benchmarks constrain evaluation.Their approaches include state-space modeling of long-range dependencies and masked-region reconstruction of enhancer–promoter coupling.Source: §5.5 · p. 6

5.6 Cross-Species Translation

  • Cross-species models align orthologous genes or conserved cell states for transfer, but incomplete ortholog mappings and domain shifts limit generalization.Sequencing depth and tissue composition are among the domain-shift sources noted.Source: §5.6 · p. 6

5.7 Generation and Simulation

  • Generative models synthesize profiles, simulate perturbations, or reconstruct modalities, but must control biological realism and avoid overfitting dominant cell types.Evaluation commonly uses distributional similarity or recovery of known cell-type proportions.Source: §5.7 · pp. 6–7

5.8 Drug-Response Prediction

  • Drug-response models map control to perturbed states and show strong predictive capacity for single-agent perturbations, while performance degrades for combinations and unseen pathways.The section identifies richer multimodal pharmacogenomic benchmarks as a need.Source: §5.8 · p. 7

6 Domain Analysis

The domain analysis scores 58 methods across ten dimensions to assess model maturity beyond task accuracy. It finds strengths in grounding, alignment, and scale alongside uneven generalization, fairness, privacy, and evaluation.

6 Domain Analysis

  • The ten-dimension rubric summarizes published performance evidence or explicit design alignment as a view of model maturity beyond task accuracy.Scores are summarized in Appendix Table 7 and Figure 2.Source: §6 · p. 7

6.1 Biological Grounding

  • Models often ground representations in marker genes, pathways, or regulatory regions, but textual and ontology grounding is inconsistent and semantics are rarely explicit in objectives.Spatial frameworks add tissue-context grounding through histology alignment.Source: §6.1 · p. 7

6.2 Batch Effects and Heterogeneity

  • Large-scale multi-domain pretraining can mitigate donor and protocol variation, but cross-platform performance still degrades on unseen sequencing chemistries or tissues.Other approaches use explicit batch tokens or adversarial alignment.Source: §6.2 · p. 7

6.3 Multi-omics Alignment

  • Cross-attention and contrastive fusion align RNA, ATAC, and protein features, while limited paired data and uneven modality coverage constrain reproducibility and benchmarking.Spatial hybrids also connect molecular and visual domains.Source: §6.3 · p. 7

6.4 Trajectory and Perturbation; 6.5 Cross-Species and Cross-Tissue Generalization

  • Dynamic models predict post-perturbation shifts with reasonable fidelity but struggle with multitarget or time-continuous responses; cross-organism transfer also drops for divergent taxa.Few methods quantify causal uncertainty or biological plausibility, and cross-species transfer is constrained by incomplete ortholog tables and human-data bias.Source: §6.4 · p. 8, §6.5 · p. 8

6.6 Atlas Fairness and Representation Balance

  • Human- and immune-cell-dominant atlases skew representation, while fairness metrics are rarely adopted and no model enforces demographic or tissue-specific parity during training.Teddy introduces benchmarking for fairness and representation balance.Source: §6.6 · p. 8

6.7 Explainability and Interpretability; 6.8 Privacy and Ethics

  • Ontology-linked text-bridge systems and agentic explanations support interpretability, but foundation-model attention visualizations offer limited causal transparency; privacy protections remain absent.The survey reports no single-cell LLM with federated or privacy-preserving learning, and ethical guidance for multi-omic sharing remains informal.Source: §6.7 · p. 8, §6.8 · p. 8

6.9 Scalability and Efficiency; 6.10 Emerging Paradigms and Agentic Behavior

  • Efficient transformers and state-space models reduce memory cost, while CellFM trains across 100 M cells; agentic systems score highest in explainability and cross-modal planning but lack standardized reasoning evaluation.High compute requirements still limit accessibility, and agentic evaluation lacks standards for reasoning fidelity or reproducibility.Source: §6.9 · p. 8, §6.10 · p. 8

7 Open Problems

The open problems center on validation, representation, multimodal integration, interpretability, privacy, and agentic reliability. The section identifies specific evidence gaps and research needs in each area.

7.1 Trust and Validation

  • Evaluation remains inconsistent, often emphasizes reconstruction over biological plausibility, and rarely includes independent replication.The section calls for community-curated leaderboards and standardized validation sets.Source: §7.1 · p. 8

7.2 Data and Bias

  • Human and mouse atlases dominate training data, leaving rare-cell, plant, and microbial systems underrepresented and limiting cross-species and clinical generalization.The text identifies diverse, balanced datasets as critical for equitable model development.Source: §7.2 · p. 8

7.3 Cross-Modal and Dynamic Modeling

  • Existing models handle pairwise modalities, while four-way RNA, ATAC, spatial, and temporal fusion remains computationally infeasible.Unified tokenization and multimodal pretraining pipelines are proposed as possible routes to consistent cross-domain reasoning.Source: §7.3 · p. 8

7.4 Interpretability and Causality

  • LLM embeddings capture statistical correlation more readily than mechanistic insight, motivating explicit reasoning layers and experimental grounding.The section identifies symbolic or causal discovery modules combined with transformers as a next step.Source: §7.4 · p. 8

7.5 Ethics and Privacy

  • Open data sharing raises privacy and dual-use concerns, while few models provide audit trails or consent-aware governance.The section calls for adapting federated and differential-privacy methods to biological data.Source: §7.5 · p. 9

7.6 Agentic and Interactive Systems

  • Agentic systems lack reliable benchmarks, and multi-step reasoning, dataset retrieval, and model orchestration remain fragile.The section calls for metrics covering reasoning accuracy, safety, and reproducibility.Source: §7.6 · p. 9

8 Conclusions

LLM4Cell synthesizes models, datasets, and evaluation domains in a unified survey of single-cell language and agentic systems. Its taxonomy and rubric frame a shift toward language-driven and increasingly autonomous analysis, with a proposed reference for future work.

8 Conclusions

  • LLM4Cell links datasets, architectures, and evaluation domains in a unified survey of 58 models and over 40 public datasets.The surveyed modalities include RNA, ATAC, multi-omic, spatial, and perturbation data.Source: §8 · p. 9
  • The synthesis describes a transition from statistical modeling toward language-driven, interpretable, and increasingly autonomous systems connecting molecular data with biological knowledge.Source: §8 · p. 9
  • The paper presents its synthesis as a reproducible reference for benchmarking, model selection, and designing next-generation cellular foundation and reasoning models.Source: §8 · p. 9

Limitations

LLM4Cell maps a 58-model corpus and 40+ public datasets, but inconsistent performance metrics, qualitative evaluation, restricted data access, and rapidly changing methods limit comparison and coverage. The survey also describes uneven dataset availability and distinct strengths and constraints across model families.

Limitations

  • Performance metrics vary across publications, the rubric lacks standardized numerical rankings, some clinical or proprietary spatial datasets are excluded, and efficiency was not systematically analyzed.The survey is a snapshot of a rapidly evolving field; non-animal resources are also scarce.Source: Limitations · p. 9

A.1 Search Strategy

  • Screening covered peer-reviewed and preprint literature from 2020–2025, using searches across PubMed, Google Scholar, arXiv, and Semantic Scholar.The search returned approximately 8,020 papers, of which 5,510 remained after filtering for English-language articles with available abstracts.Source: §A.1 · p. 14

A.2 Inclusion and Screening Criteria

  • The review included models meeting at least one of four criteria, including pretrained models with ≥10M parameters, textual grounding, agentic reasoning, or relevance to cellular foundation-model research.After merging duplicates, curation yielded 58 models across five families; preprints were included to capture recent work and manually checked for active code or data links.Source: §A.2 · p. 14, §A.3 · p. 14

A.4 Resulting Corpus

  • The corpus spans 58 methods, 37 institutions, 40+ public datasets, and eight analytical tasks, with RNA atlases larger in scale than fragmented chromatin and spatial resources.Paired modalities offer potential benchmarks, while privacy and licensing restrict clinical data and cross-species coverage remains limited.Source: §A.4 · p. 14, §B.7 · p. 15

B.1 RNA Atlases

  • The dataset landscape includes RNA atlases, chromatin and joint RNA–ATAC resources, paired multi-omic datasets, spatial maps and imaging, perturbation screens, and plant atlases.The supplied passages identify sparsity as a constraint on large-scale pretraining for chromatin data and note that the Multiome Benchmark Pack is smaller than large RNA atlases.Source: §B.1 · p. 14, §B.2 · p. 14, §B.3 · p. 14

C.1 Foundation Models (Extended Details)

  • Foundation embeddings transfer to annotation, integration, and trajectory inference but lack explicit biological grounding; ontology alignment and reasoning remain external, and interpretability is largely post-hoc.Text-Bridge systems connect molecular data with ontology labels, pathways, or literature, but remain limited by vocabulary coverage and incomplete multi-omic grounding.Source: §C.1 · p. 15, §C.2 · p. 15

C.3 Spatial and Multimodal Models (Extended Details)

  • Spatial and multimodal models combine molecular profiles with positional or image information; they show high biological grounding but moderate explainability because visual features are opaque.Evaluation includes spot-level reconstruction and correlation with histological segmentation; multi-omic extensions integrate RNA, ATAC, and protein in a shared latent space.Source: §C.3 · pp. 15–16

C.4 Epigenomic Models (Extended Details)

  • Epigenomic models show strong biological grounding and multi-omics alignment but limited interpretability, while agentic frameworks face open challenges in reasoning accuracy, privacy, and reproducibility.Agentic systems combine molecular encoders, LLM controllers, and external tools for multi-step tasks.Source: §C.4 · p. 16, §C.5 · p. 16

D Datasets

The appendix tables inventory single-cell datasets by modality and research area.

RNA-seq datasets

  • Table 1 inventories RNA-seq single-cell datasets used in LLM-based research.Compare the listed resources as an RNA-seq dataset inventory.Source: Table 1 · p. 17

ATAC-seq datasets

  • Table 2 inventories ATAC-seq and chromatin-accessibility datasets used in LLM-based research.Compare the listed resources within this chromatin-data scope.Source: Table 2 · p. 18

Multiome datasets

  • Table 3 inventories paired and tri-omic datasets, including RNA–ADT, RNA–ATAC, ATAC–ADT, and tri-modal combinations.Compare the listed resources by modality combination.Source: Table 3 · p. 19

Spatial datasets

  • Table 4 inventories spatial transcriptomics and imaging datasets used in single-cell research.Compare the listed resources within the spatial and imaging scope.Source: Table 4 · p. 20

Perturbation and drug-response datasets

  • Table 5 inventories perturbation and drug-response single-cell datasets used in research frameworks.Compare the listed resources within the perturbation and drug-response scope.Source: Table 5 · p. 21

Plant datasets

  • Table 6 inventories plant RNA single-cell transcriptomics datasets used in LLM research.Compare the listed resources within the plant transcriptomics scope.Source: Table 6 · p. 22

E Methods Comparison

The methods comparison covers atlas-pretrained, generative, cross-species, multimodal, spatial, text-bridging, epigenomic, and agentic approaches, with described uses ranging from annotation and perturbation to imputation, prediction, and spatial-region matching. Appendix figures contrast task and domain coverage between agentic and non-agentic models.

E Methods Comparison

  • scFoundation and CellFM learn atlas-scale cell representations for annotation and perturbation, while iSEEEK uses gene-ranking similarity for scalable cross-dataset embeddings and clustering.The comparison describes scFoundation as using a large pretrained corpus and CellFM as learning embeddings from 100M cells.Source: E Methods Comparison · p. 23
  • tGPT treats gene-expression ranks as token sequences for generative embeddings in clustering and trajectory inference, while scBERT and UCE support annotation and cross-species cell analysis.scBERT is described as enabling novel cell discovery, and UCE embeds cells zero-shot in a shared latent space.Source: E Methods Comparison · p. 23
  • GeneCompass learns cross-species gene–cell embeddings for annotation, perturbation, dose-response analysis, and gene-regulatory-network inference.Its listed ontology incorporates gene-regulatory networks, coexpression, and gene-family knowledge.Source: E Methods Comparison · p. 23
  • Multimodal methods cover RNA-plus-protein and transcriptomics-plus-proteomics data, using modality-specific transformers, cross-attention, or transformer fusion for imputation and label transfer.The comparison also lists classification, representation fusion, and recovery of missing modalities among their functions.Source: E Methods Comparison · p. 24
  • scMulan combines text-derived gene metadata embeddings with expression data to support clustering, batch correction, annotation, and perturbation.The comparison characterizes its embeddings as LLM-generated from gene or metadata descriptions plus raw data.Source: E Methods Comparison · p. 24
  • Spatial approaches extend scGPT through spatial-aware decoding for slide integration, deconvolution, and gene imputation, while hierarchical graph modeling supports spatial cell-and-gene embeddings and clinical outcome prediction.The comparison also lists cell annotation and spatial clustering among the graph-based approach’s tasks.Source: E Methods Comparison · p. 26
  • Other spatial methods contextualize gene embeddings with ligand signals or fuse histology and spot features to predict spatial expression, denoise data, and detect heterogeneity.The comparison also describes unified visual-omics embeddings for cross-modal prediction, annotation, and tissue retrieval.Source: E Methods Comparison · p. 26
  • A language-oriented spatial method turns spatial-transcriptomics data and regions of interest into narratives, then matches natural-language queries to regions through GO-term and LLM interpretation.The comparison separately lists SpaCCC and spaLLM among spatial-transcriptomics methods.Source: E Methods Comparison · pp. 26–27
  • Spatial approaches embed ligand–receptor pairs in LLM latent space for communication inference and fuse LLM embeddings with spatial-omics signals through multi-view attention for domain identification.The communication workflow uses diffusion and permutation-test filtering.Source: E Methods Comparison · p. 27
  • Text-bridging methods link single-cell and text language models for cell descriptions, text-generated pseudo-cells, and annotation, or encode cells as gene-ranked sentences for biological-text generation and reconstruction.A related approach trains large LLMs on cell sentences and biological text for perturbation prediction and multicellular summarization.Source: E Methods Comparison · p. 27
  • Epigenomic modeling learns regulatory syntax from chromatin accessibility and DNA motifs to predict expression and interpret transcription-factor–cis interactions.The comparison identifies scATAC as the chromatin-accessibility modality and EpiFoundation as an epigenomic foundation model.Source: E Methods Comparison · pp. 27–28
  • Gene-text embeddings can be aggregated by gene weights or ranks into cell embeddings, while contrastive-learning transformers provide cell representations for annotation and drug-sensitivity prediction.The comparison also identifies a contrastive embedding learned from expression data.Source: E Methods Comparison · p. 28
  • scExtract uses LLMs to extract pipeline parameters from publications and apply them to dataset harmonization.Its comparison describes article-based parameter extraction for annotation and integration.Source: E Methods Comparison · p. 28

F Appendix Figure

  • Appendix figures show agentic models emphasizing annotation, ontology mapping, spatial mapping, explainability, fairness, and emerging paradigms, while non-agentic models emphasize trajectory, perturbation, regulatory and pathway inference, biological grounding, and batch effects.The task-coverage comparison is shown in Figure 4 and the domain-coverage comparison in Figure 5.Source: Figure 4 · p. 35, Figure 5 · p. 35
Loading 2510.07793v3…