Title: CodeCytos: Leveraging Code-Augmented Agent Action Spaces for AI-Driven Spatial Molecular Imaging Analysis

Conventional software for tissue image analysis typically offers essential functions such as cellular segmentation, basic morphological feature extraction, and spatial organization assessment. Nevertheless, these standard tools frequently demand significant manual input and lack seamless integration with code-based automation, thereby constraining their efficiency and scalability for intricate spatial tissue investigations. Furthermore, their utility for bespoke analyses is restricted, as they generally accommodate only a predefined repertoire of spatial cellular features.

To overcome these challenges, we introduce CodeCytos, a framework built on coding-based reasoning agents that facilitates dynamic, programmable engagement with spatial molecular imaging data. This approach aims to enhance both automation and customization capabilities. CodeCytos is specifically engineered to simplify the investigation of custom spatial cellular features, allowing it to adjust to a wide array of research requirements.

We validated the framework’s effectiveness through case studies involving four expertly curated datasets representing distinct tissue types: the frontal cortex, non-small-cell lung cancer, pancreas, and tonsil. Our evaluation employed a realistic minimal prompt scenario, in which bioscientists asked straightforward questions without providing task-specific directives or context regarding spatial cellular analysis. During this process, we benchmarked various large language model (LLM) backbones known for robust coding proficiency.

Additionally, we demonstrated that integrating tailored, domain-agnostic few-shot in-context coding-reasoning examples—specifically, randomly selected demonstrations from outside the spatial analysis field—can significantly boost performance. This improvement occurs without the need for expensive, expert-designed demonstrations within the specific domain. Ultimately, CodeCytos surpasses baseline methods, underscoring the promise of code-action agents in supporting custom feature exploration within spatial molecular imaging and expediting the discovery of biomarkers.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC