TLDR: LLM4Cell is the first comprehensive survey of 58 large language and agentic models for single-cell biology. It categorizes these models into five families (foundation, text-bridge, spatial/multimodal, epigenomic, and agentic), maps them to eight key analytical tasks, and evaluates them across ten domain dimensions using over 40 public datasets. The survey provides an integrated view of language-driven single-cell intelligence, highlighting current progress, existing fragmentation, and outlining open challenges in interpretability, standardization, and trustworthy model development.
The field of single-cell biology is experiencing a significant transformation with the advent of large language models (LLMs) and agentic frameworks. These advanced AI tools are beginning to enable natural-language reasoning, generative annotation, and the integration of diverse data types, offering new avenues for understanding cellular processes. However, the rapid progress in this area has also led to a fragmented landscape, with models varying widely across data modalities, architectural designs, and evaluation standards.
To address this growing complexity, a new comprehensive survey, LLM4Cell, has been introduced. This pioneering work provides the first unified overview of 58 foundational and agentic models specifically developed for single-cell research. It spans various data types, including RNA, ATAC, multi-omic, and spatial modalities, offering a much-needed framework for understanding this evolving domain. For more details, you can refer to the full research paper.
Categorizing the AI Landscape
LLM4Cell categorizes these diverse methods into five distinct families: foundation, text-bridge, spatial/multimodal, epigenomic, and agentic models. This classification helps in mapping their capabilities to eight key analytical tasks crucial in single-cell biology. These tasks include cell annotation, modeling cell trajectories and responses to perturbations, and predicting drug responses. The survey highlights a clear progression from models that learn basic representations from single-cell data to more sophisticated systems capable of reasoning, dialogue, and autonomous analysis.
Data and Evaluation
The survey draws on over 40 publicly available datasets, covering a wide range of modalities such as RNA, ATAC, multi-omic, spatial, perturbation, and even plant single-cell data. This extensive collection allows for a thorough analysis of benchmark suitability, data diversity, and potential ethical or scalability constraints. Furthermore, LLM4Cell evaluates models across ten critical domain dimensions. These dimensions encompass aspects like biological grounding, how well models align with multi-omics data, fairness, privacy considerations, and explainability, providing a holistic view of model maturity and trustworthiness.
Key Model Families and Their Roles
Foundation Models: These are the bedrock, learning transferable cell and gene representations directly from large-scale single-cell RNA sequencing data without explicit labels. Examples include scGPT and Geneformer, which are crucial for tasks like annotation and integration.
Text-Bridge LLMs: These models connect molecular data with biomedical language, grounding single-cell representations in semantic and ontological knowledge. They enhance interpretability and enable zero-shot annotation by aligning gene or cell embeddings with textual descriptions.
Spatial and Multimodal Models: These frameworks integrate gene expression with spatial coordinates, histology, or additional omics data to capture the intricate architecture of tissues. Models like TransformerST and OmiCLIP are vital for spatial mapping and understanding cellular context within tissues.
Epigenomic Models: Extending LLM concepts to chromatin accessibility and regulatory data, these models learn cis-regulatory patterns and infer gene-regulatory networks, improving biological grounding in epigenetics.
Agentic Frameworks: Representing the cutting edge, agentic systems combine pretrained models with reasoning modules for autonomous single-cell analysis. They can plan tasks, query ontologies, and interact with tools, enabling dialogue-based annotation and multi-step reasoning.
Also Read:
- MoRE-GNN: A Data-Driven Graph Neural Network for Multi-omics Single-Cell Integration
- Assessing Multimodal AI Retrieval in Medical Applications
Addressing Open Challenges
Despite significant advancements, the LLM4Cell survey identifies several open challenges. These include the need for more consistent evaluation metrics, particularly for reasoning correctness and biological plausibility. Data scarcity and bias, especially in non-human and clinical spatial datasets, remain significant hurdles. Achieving true cross-modal integration and developing models with better interpretability and causal reasoning are also critical. Furthermore, ethical considerations, such as data privacy and the development of trustworthy AI for cell biology, are highlighted as essential areas for future research.
By linking datasets, models, and evaluation domains, LLM4Cell offers an integrated perspective on language-driven single-cell intelligence. It serves as a crucial reference for benchmarking, model selection, and guiding the design of next-generation cellular foundation and reasoning models, paving the way for a more unified and interpretable understanding of single-cell biology.
