SP-Mind autonomous agent for spatial proteomics, SP-Bench 102

SP-Mind, an autonomous AI agent for spatial proteomics analysis, unifies the pipeline from raw multiplexed tissue imaging to downstream phenotype discovery. The paper was submitted to arXiv on 23 Jun 2026 (arXiv:2606.24235), is 23 pages with 6 figures, and the authors note acceptance to ICML 2026.

What is SP-Mind?

SP-Mind is the first autonomous reasoning agent explicitly built to run end-to-end spatial proteomics analysis, converting natural-language queries into analytical workflows without task-specific fine-tuning. The system bundles expert-curated biological analysis skills with specialized computational tools to chain together operations from image-level inputs through phenotype discovery.

The authors frame spatial proteomics as a single-cell-resolution approach that profiles protein expression inside tissue architecture. They position SP-Mind as a unifying layer over fragmented workflows that today require manual orchestration of heterogeneous tools. The paper lists Yucheng Yuan and Yuanfeng Ji as equal contributors, with coauthors Zhongxiao Li and Ruijiang Li.

How was SP-Mind evaluated and what is SP-Bench?

SP-Mind was evaluated on a new benchmark called SP-Bench, which the paper describes as spanning diverse tissue types and comprising 102 tasks across 18 distinct categories. The authors ran extensive experiments on SP-Bench and on established downstream tasks and report that SP-Mind achieves state-of-the-art performance compared to existing open-source biomedical agent baselines.

SP-Bench is presented as a comprehensive test suite designed to exercise the end-to-end pipeline from raw multiplexed tissue imaging to phenotype-level outputs. The paper does not publish numeric performance breakdowns in the abstract, but it emphasizes scale and breadth by naming the 102-task, 18-category structure as a central evaluation artifact.

How does SP-Mind work at a high level?

SP-Mind synthesizes domain expertise and computational modules to translate natural-language requests into multi-step analysis plans and then executes them with the appropriate tools. The agent pairs expert-curated biological analysis skills with specialized computation modules to perform tasks ranging from image preprocessing to phenotype discovery.

The manuscript treats SP-Mind as a coordinator: it accepts a natural-language query, selects and composes skills and tools, and produces analysis outputs without per-task fine-tuning. The authors highlight this capability as a response to fragmented, manually orchestrated workflows in spatial proteomics.

Why it matters

SP-Mind addresses a bottleneck in spatial proteomics: fragmented pipelines that demand manual assembly of diverse tools and expert oversight. By automating plan generation and tool orchestration from natural-language inputs, SP-Mind could make spatial-proteomics workflows more reproducible and scalable, lowering the barrier for researchers to run complex end-to-end analyses.

The inclusion of SP-Bench as a 102-task, 18-category benchmark gives the project a measurable, reproducible evaluation target. If the community adopts SP-Bench, comparisons across agents and toolchains will become more standardized, helping groups judge progress on pipeline automation rather than isolated model components.

What to watch

Look for the full paper and supplementary materials linked from arXiv:2606.24235, which should contain detailed per-task results, examples of the agent-generated workflows, and the set of expert-curated skills and tools. Acceptance to ICML 2026 means the community will soon see reviewers' context and possibly a conference presentation that clarifies limitations and failure modes.

The authorship and submission metadata: submitted 23 Jun 2026, authors Yucheng Yuan, Yuanfeng Ji, Zhongxiao Li, and Ruijiang Li; paper length 23 pages with 6 figures. The DOI listed on arXiv is https://doi.org/10.48550/arXiv.2606.24235.