SPARKing New Insights: Agentic AI Framework Validates Novel Biomarkers Across Five Cancer Types

Investigators have created a foundational agentic AI approach, called SPARK (System of Pathology Agents for Research and Knowledge), for autonomously generated biologically driven tumor analysis using complex pathology data and analytical tools. Explanation of the framework was published in Nature Medicine. Through the SPARK framework, which the study authors said functions as a “pathology brain,” the researchers gained new biological insights into tumor evolution.

“SPARK helps to refine diagnoses, stratify patients more reliably, and make more precise treatment decisions. Particularly in the field of personalized oncology, there is an opportunity to tailor treatments more closely to the individual biological characteristics of a tumor, thereby improving treatment outcomes,” stated senior author Yuri Tolkach, MD, PhD, Senior Physician, Institute of Pathology, University Hospital Cologne and Medical Faculty, University of Cologne.

Model Methods

SPARK consists of a set of interconnected AI agents that use language as a universal interface for more efficient interaction with complex data from routine hematoxylin and eosin (H&E)–stained whole-slide images. Large language models were carefully selected for each agent focused on four main tasks. The agents autonomously generate biological hypotheses, refine these ideas, and code and verify the idea and parameters with analytical tools to assess the validity of the hypotheses. A dedicated module was also created for human interaction with SPARK for a human-in-the-loop option.

The investigators also noted that the modular agentic workflow does not require additional model training or retraining when working with new biological concepts and analytical tools.

They tested the framework on 18 cohorts of patients with resected malignant tumors across five cancer types, including lung adenocarcinoma, lung squamous cell carcinoma, colorectal cancer, breast cancer, and oropharyngeal squamous cell carcinoma. All patients had H&E tumor slides plus clinical prognostic and predictive data and had not yet received any neoadjuvant therapy. These patient cohorts were gathered from open-source data sets, including The Cancer Genome Atlas and others.

Additionally, the framework was evaluated a spatial biology dataset of 625 patients with well-characterized breast cancer.

Findings

The SPARK framework produced relevant clinical and biological concepts that were related to known biomarkers, such as patterns of tumor progression and temporal change, and correlated with prognosis and treatment response.

In the Nature Medicine report, the study authors presented three use case examples for SPARK based on the test cohorts.

The first focused on prognostic tissue biomarkers from the H&E slides to make patient survival predictions. SPARK generated 500 unique ideas that were compiled into 1,115 verified parameters, some correlating with known biomarkers.

The second use case looked at ways to determine if a malignant tumor has metastasized. The framework generated 118 ideas, but no one parameter could capture the full complexity of metastatic behavior.

And the third use case assessed multiplexing and spatial biology data to extract independent prognostic markers from the breast cancer cohort. SPARK generated 561 ideas and 2,452 verified parameters. After making adjustments for grade, stage, and receptor status, many parameters showed independent prognostic value.

“With SPARK, we aim to transform pathology from a primarily descriptive discipline into a data-driven, predictive science—and thereby make a significant contribution to precision medicine in oncology,” said study author Reinhard Büttner, MD, Director of the Institute of General Pathology and Pathological Anatomy, University of Cologne.

The study authors noted, however, that further prospective validation is needed to assess the clinical utility of the framework.

DISCLOSURES: SPARK was funded, amongst others, by the former German Federal Ministry of Education and Research (BMFTR) and as part of the DigiPathConnect project under the European Union’s Interreg Euregio Meuse-Rhine program. Open access funding was provided by Universität zu Köln. The authors declared no competing interests. For access to the open-source code, parameters, and data used in the agentic pipeline, visit nature.com.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.