Empowering biomedical evidence exploration and synthesis with deep knowledge graph research

Area of Science:

• Biomedical informatics
• Artificial intelligence in medicine
• Knowledge representation

Background:

• Generic artificial intelligence (AI) agents have limitations in complex biomedical research.
• There is a need for advanced AI systems capable of deep literature exploration and evidence synthesis.

Purpose of the Study:

• To introduce DeepEvidence, a deep research agent designed for comprehensive evidence exploration and synthesis.
• To evaluate DeepEvidence's performance against generic AI agents and on benchmark tasks across the biomedical research spectrum.

Main Methods:

• DeepEvidence employs coordinated multi-agent collaboration using breadth-first and depth-first strategies.
• It integrates evidence from heterogeneous biomedical knowledge sources and literature.
• An evidence graph is incrementally constructed for transparent tracking and validation.

Main Results:

• DeepEvidence significantly outperforms generic AI agents on four open benchmarks.
• Demonstrated substantial improvements in systematic evidence exploration and synthesis across seven new benchmark tasks.
• Showcased effectiveness in areas including drug discovery, preclinical studies, clinical trials, and evidence-based medicine.

Conclusions:

• Deep research agents, exemplified by DeepEvidence, offer advanced capabilities for biomedical research.
• DeepEvidence has the potential to accelerate biomedical discovery and translational research through sophisticated evidence synthesis.