A new AI system called SpatialAgent has been developed to simplify spatial biology workflows, including tasks like gene panel design and hypothesis generation. Created by a collaboration between Genentech, Stanford, and several academic institutions, the system integrates large language models (LLMs) with domain-specific tools and adaptive reasoning strategies.
SpatialAgent is designed to function autonomously or in collaboration with scientists. It supports complex spatial transcriptomics tasks through an internal framework composed of memory, planning, and action modules. These modules coordinate the selection and execution of appropriate tools while maintaining both short-term and long-term task context. The system dynamically adapts to user feedback, new data types, and biological contexts without requiring rigid preprogrammed workflows.
The researchers benchmarked SpatialAgent across multiple tasks. For gene panel design, it was tested on human dorsolateral prefrontal cortex data and outperformed four established computational methods and 90% of human experts in cell-type prediction. When paired with human-designed panels, SpatialAgent improved performance in most cases and occasionally surpassed its autonomous output.
The tool also showed strong performance in annotating cell types and tissue niches in developing human heart tissue. It matched or outperformed expert annotations and outclassed methods like CellTypist and GPTCellType, particularly in multimodal datasets. SpatialAgent’s ability to incorporate histology and expression data enabled more consistent and spatially coherent annotations across samples.
In a mouse colitis model, SpatialAgent independently identified known and additional cell-cell communication patterns using integrated tools such as LIANA and CellPhoneDB. Its findings were consistent with a recent study and expanded on it by highlighting fibroblast–pericyte interactions and signaling pathways not previously emphasized.
Finally, SpatialAgent was applied in a real-world setting to augment a standard spatial transcriptomics panel for prostate cancer research. The resulting hybrid panel improved cell type resolution and revealed distinct interaction networks in the tumor microenvironment.
The work demonstrates that autonomous AI agents can effectively support spatial biology research, providing interpretable outputs and reducing reliance on expert-driven workflows. The SpatialAgent code is available at https://github.com/Genentech/SpatialAgent.
Publication Details
Hanchen Wang et al. SpatialAgent: An autonomous AI agent for spatial biology. bioRxiv. 2025.04.03.646459; doi: https://doi.org/10.1101/2025.04.03.646459
SpatialAgent is designed to function autonomously or in collaboration with scientists. It supports complex spatial transcriptomics tasks through an internal framework composed of memory, planning, and action modules. These modules coordinate the selection and execution of appropriate tools while maintaining both short-term and long-term task context. The system dynamically adapts to user feedback, new data types, and biological contexts without requiring rigid preprogrammed workflows.
The researchers benchmarked SpatialAgent across multiple tasks. For gene panel design, it was tested on human dorsolateral prefrontal cortex data and outperformed four established computational methods and 90% of human experts in cell-type prediction. When paired with human-designed panels, SpatialAgent improved performance in most cases and occasionally surpassed its autonomous output.
The tool also showed strong performance in annotating cell types and tissue niches in developing human heart tissue. It matched or outperformed expert annotations and outclassed methods like CellTypist and GPTCellType, particularly in multimodal datasets. SpatialAgent’s ability to incorporate histology and expression data enabled more consistent and spatially coherent annotations across samples.
In a mouse colitis model, SpatialAgent independently identified known and additional cell-cell communication patterns using integrated tools such as LIANA and CellPhoneDB. Its findings were consistent with a recent study and expanded on it by highlighting fibroblast–pericyte interactions and signaling pathways not previously emphasized.
Finally, SpatialAgent was applied in a real-world setting to augment a standard spatial transcriptomics panel for prostate cancer research. The resulting hybrid panel improved cell type resolution and revealed distinct interaction networks in the tumor microenvironment.
The work demonstrates that autonomous AI agents can effectively support spatial biology research, providing interpretable outputs and reducing reliance on expert-driven workflows. The SpatialAgent code is available at https://github.com/Genentech/SpatialAgent.
Publication Details
Hanchen Wang et al. SpatialAgent: An autonomous AI agent for spatial biology. bioRxiv. 2025.04.03.646459; doi: https://doi.org/10.1101/2025.04.03.646459