SPARK
System of Pathology Agents for Research and Knowledge — an agentic framework that autonomously generates biologically driven concepts, codes them into analytical tools, and applies them to H&E and spatial-biology data across five cancer types.
| Affiliation | University Hospital Cologne, Martin Luther University Halle-Wittenberg, University Hospital Essen, University Hospital Giessen, and collaborators (corresponding: Yuri Tolkach) |
| First introduced | 2026-03 (accepted Nature Medicine; received 2025-09-02, accepted 2026-03-19) |
| Lifecycle stages | Multi-stage (idea generation, idea refinement, idea/parameter coding, parameter verification) |
| Autonomy level | Fully autonomous in the discovery loop, with an interactive module for clinician-driven queries |
| Domain focus | Computational cancer pathology — H&E whole-slide imaging and multiplexed spatial biology |
| Availability | Open source — code, parameters and results released (GitHub) |
Approach
Four linked modules use language as the universal interface between agents: an Idea Generation Agent (OpenAI o1 reasoning model, Jan–Feb 2025) proposes tumor-analysis concepts under varying creativity levels; an Idea Refinement Agent and Duplicate Detector filter and consolidate; an Idea Coding Agent (Claude Sonnet) converts concepts into executable Python parameters with up to three implementation attempts and Code Review feedback; and a Parameter Verification Agent screens for biases and instability. SPARK ingests routine H&E whole-slide images preprocessed by tissue segmentation and single-cell detection/classification across seven cell types (tumor, fibroblasts, macrophages, lymphocytes, neutrophils, eosinophils, plasma cells), then runs the agent-generated parameters case-wide and aggregates per-slide outputs to case level.
Validation
Evaluated across 18 patient cohorts and more than 5,400 patients in five cancer types — lung adenocarcinoma, lung squamous cell carcinoma, colorectal, breast, and oropharyngeal squamous cell carcinoma — in both prognostic and predictive settings, plus a multiplexed spatial-biology breast-cancer dataset (n = 625). Parameters were correlated against histologic grade, pN stage, established biomarkers, and clinical follow-up.
Notable results
- 99.2% of LLM-proposed parameters compiled to executable code; after quality filtering, 475 ideas yielded 1,115 non-redundant parameters from an initial 2,368 coded.
- Discovered parameters correlated with prognosis and predictive biomarkers across all five tumor types, and inferred patterns of tumor progression and temporal change from static H&E images.
- Open-source LLM alternatives (general and medical-domain) were shown to be viable substitutes for both idea generation and parameter coding.
Primary paper
Other references
None yet.
Code
github.com/cpath-ukk/SPARK — full agentic pipeline and coded parameter library.