Transplant-Agents : A Multi-Agent Artificial Intelligence Framework for Reproducibility Assessment of Post-Transplant...

Area of Science:

• Biomedical Informatics
• Artificial Intelligence in Medicine
• Transplantation Science

Background:

• Reproducible biomarker identification and transplant rejection risk prediction are critical unmet needs in transplantation.
• Current methods are often hypothesis-driven, limiting scalability and generalizability.

Purpose of the Study:

• To introduce Transplant-Agents, a data-driven multi-agent AI framework for automated biomarker discovery and transplant rejection risk prediction.
• To demonstrate the framework's ability to achieve reproducible results across multiple iterations.

Main Methods:

• Integration of large language models (LLMs) with machine learning algorithms within a multi-agent AI system.
• Agents engage in structured, iterative dialogue governed by predefined rules for biomarker selection.
• Evaluation on three multicenter clinical trial transplant datasets (kidney, liver, heart) from ImmPort, including 683 patients.

Main Results:

• Transplant-Agents achieved high predictive performance with AUROC scores of 0.93 (kidney), 0.88 (liver), and 0.88 (heart).
• Feature importance analysis confirmed the stability, interpretability, and generalizability of the identified biomarkers.
• The framework successfully reproduced established transplant biomarkers.

Conclusions:

• AI-agent frameworks, like Transplant-Agents, offer a reliable method for biomarker identification and transplant rejection risk prediction.
• This approach enables transparent, validated, and standardized risk prediction pipelines in transplantation medicine.