SP-Mind: An Autonomous Reasoning Agent for Spatial Proteomics Analysis

Research Objective

Spatial proteomics is essential for understanding tumor microenvironments, yet the analysis pipeline is fragmented, requiring manual orchestration of diverse, complex tools. The authors sought to address this by developing SP-Mind, an autonomous AI agent capable of interpreting natural-language queries to execute end-to-end spatial proteomics workflows without the need for task-specific fine-tuning.

SP-Mind utilizes a ReAct-style reasoning loop, which allows the agent to observe data, think through necessary steps, and execute code. The agent is equipped with two primary innovations: a modular library of specialized computational tools (covering image preprocessing, registration, segmentation, and quantification) and 'Spatial BioSkill Templates.' These templates are expert-curated knowledge primitives that provide the agent with procedural guidance, parameter heuristics, and error-recovery protocols. To validate the system, the authors introduced SP-Bench, a new benchmark consisting of 102 tasks across 18 categories, ranging from basic single-stage operations to complex, multi-stage end-to-end analyses.

SP-Mind demonstrated state-of-the-art performance on the SP-Bench framework, achieving 68.9% execution accuracy and outperforming existing biomedical agent baselines by 13 percentage points. The results indicate that the integration of domain-specific skill templates significantly enhances the agent's ability to handle the multi-step, dependency-heavy nature of spatial biology workflows compared to general-purpose agentic frameworks.

By automating the orchestration of complex spatial proteomics pipelines, SP-Mind reduces the barrier to entry for researchers, improves reproducibility, and enables more scalable analysis of multiplexed tissue imaging. This work represents a significant step toward 'AI Scientists' that can handle the full lifecycle of high-dimensional biological data analysis.