An agentic framework for autonomous scientific discovery in cancer pathology

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Biomedical informatics

Background:

• Current AI in cancer pathology often relies on manual features, lacks explainability, and uses fragmented workflows.
• There is a need for AI systems that can autonomously generate biologically relevant insights directly from complex pathology data.

Purpose of the Study:

• To introduce SPARK (System of Pathology Agents for Research and Knowledge), an agentic AI approach using language as an interface for autonomous tumor analysis.
• To evaluate SPARK's ability to generate clinically and biologically relevant concepts from histopathology images across diverse cancer types.

Main Methods:

• SPARK utilizes a language-based interface to autonomously translate biological concepts into analytical tools for pathology data.
• The system was evaluated on 18 patient cohorts (over 5,400 patients) across five cancer types (lung adenocarcinoma, lung squamous cell carcinoma, colorectal, breast, and oropharyngeal squamous cell carcinoma).
• Evaluation included prognostic and predictive settings, as well as a spatial biology dataset for breast cancer (625 patients).

Main Results:

• SPARK generated clinically and biologically relevant concepts correlated with patient prognosis.
• The generated concepts showed correlations with known pathological variables and predictive biomarkers.
• SPARK inferred patterns of tumor progression and temporal changes from static histopathology images.

Conclusions:

• SPARK demonstrates potential for autonomous generation of biologically driven insights in cancer pathology.
• The system offers a novel approach to analyzing complex pathology data, potentially improving diagnostic precision and deepening tumor biology understanding.
• Open release of code and results aims to facilitate further research and clinical application, though prospective validation is required.