Abstract
Biomedical knowledge graphs (KGs) encode vast, heterogeneous information spanning literature, genes, pathways, drugs, diseases, and clinical trials, but leveraging them collectively for scientific discovery remains difficult. Their structural differences, continual evolution, and limited cross-resource alignment require substantial manual integration, limiting the depth and scale of knowledge exploration. We introduce DeepEvidence, an AI-agent framework designed to perform Deep Research across various heterogeneous biomedical KGs. Unlike generic Deep Research systems that rely primarily on internet-scale text, DeepEvidence incorporates specialized knowledge-graph tooling and coordinated exploration strategies to systematically bridge heterogeneous resources. At its core is an orchestrator that directs two complementary agents: Breadth-First ReSearch (BFRS) for broad, multi-graph entity search, and Depth-First ReSearch (DFRS) for multi-hop, evidence-focused reasoning. An internal, incrementally built evidence graph provides a structured record of retrieved entities, relations, and supporting evidence. To operate at scale, DeepEvidence includes unified interfaces for querying diverse biomedical APIs and an execution sandbox that enables programmatic data retrieval, extraction, and analysis. Across established deep-reasoning benchmarks and four key stages of the biomedical discovery lifecycle: drug discovery, pre-clinical experimentation, clinical trial development, and evidence-based medicine, DeepEvidence demonstrates substantial gains in systematic exploration and evidence synthesis. These results highlight the potential of knowledge-graph-driven Deep Research to accelerate biomedical discovery.
Introduction
The biomedical domain is characterized by an explosion of data distributed across specialized Knowledge Graphs (KGs) such as KEGG for pathways, ChEMBL for chemical data, and OpenTargets for gene-disease associations. While these resources are critical for discovery, they remain highly fragmented and evolve independently. Current "Deep Research" AI agents typically rely on internet-scale unstructured text, which often lacks the precision and multi-hop reasoning required for complex scientific tasks—such as linking specific genetic mutations to metabolic flux responses. DeepEvidence addresses this bottleneck by enabling AI agents to navigate structured databases systematically, bridging the gap between raw data and mechanistic insight.
Methodology
DeepEvidence utilizes a hierarchical multi-agent architecture designed to move beyond simple text retrieval toward structured graph traversal.
Multi-Agent Orchestration
The system is governed by an Orchestrator that decomposes complex queries into actionable plans. These plans are executed by two specialized sub-agents:
- Breadth-First ReSearch (BFRS): This agent conducts wide-scale surveys across multiple KGs to identify "seed" entities and establish broad relationships.
- Depth-First ReSearch (DFRS): This agent follows specific evidence chains, such as citation trails or metabolic pathways, to uncover detailed mechanistic insights.
Internal Evidence Graph and Execution Sandbox
Unlike standard Retrieval-Augmented Generation (RAG) systems that utilize flat document lists, DeepEvidence maintains an Internal Evidence Graph. This structured memory tracks entities, relations, and their provenance, ensuring that every conclusion is grounded in verifiable facts. To interact with the data, agents operate within an Execution Sandbox, writing and executing Python code to query over 17 biomedical APIs (e.g., PubMed, UniProt, ClinicalTrials.gov). This programmatic approach allows for sophisticated data aggregation and statistical analysis, mirroring the workflow of a human bioinformatician.
Evaluation and Results
The framework was evaluated on both established reasoning benchmarks and a newly curated suite of tasks covering the drug discovery lifecycle, including target identification, clinical trial design, and evidence gap discovery.
DeepEvidence consistently outperformed frontier models (such as GPT-5 and Claude 3.5 Sonnet) and specialized biomedical agents (Biomni, ToolUniverse):
- HLE-Medicine: Achieved 40.0% accuracy, a massive jump over GPT-5’s 3.3%.
- LabBench-LitQA2: Reached 80.0%, more than doubling the performance of specialized agents like Biomni (32.0%).
- Clinical Trial Planning: In sample size estimation tasks, it reached 68% accuracy, significantly higher than the 20-32% seen in baselines.
- Evidence-Based Medicine: In detecting gaps in clinical trial reviews, it achieved a 90% gap detection rate, compared to just 10-50% for other systems.
Discussion and Future Work
DeepEvidence demonstrates the power of agentic traversal over fixed algorithmic rules, allowing for the discovery of "bridge entities" across heterogeneous databases. However, limitations remain. While the system queries 17 major KGs, many proprietary or niche databases are not yet integrated. Furthermore, while DeepEvidence surpasses current AI baselines, it has not yet reached the level of senior human experts in highly open-ended research tasks.
Future developments will focus on incorporating multimodal data—such as pathology slides, 3D protein structures, and proteomics—and enabling agents to autonomously discover and build tools for unfamiliar databases. This evolution aims to bring AI-driven research closer to the complexity and intuition of expert human discovery.
Summary updated