DeepEvidence: Empowering Biomedical Discovery with Deep Knowledge Graph Research

TL;DR

DeepEvidence is introduced, an AI-agent framework designed to perform Deep Research across various heterogeneous biomedical KGs, and demonstrates substantial gains in systematic exploration and evidence synthesis.

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.

Figures (7)

• Figure 1: Overview of the DeepEvidence and the benchmark performance. (a) System overview showing the orchestrator coordinating planning, action, and memory updates across research agents. A user query is translated into a planned sequence of research agents that act within an execution sandbox and update a shared memory through observations. These agents perform coordinated breadth-first and depth-first exploration across multiple biomedical knowledge graphs to identify cross-graph links and build an integrated evidence view. (b) The agent incrementally builds a structured graph that links papers, drugs, genes, and pathways through relations such as targets, tests, and citations. This evidence graph serves as a persistent memory that organizes discovered entities and supports downstream reasoning. (c) Overview of the diverse biomedical knowledge graphs that the agent can access across domains, including genes, diseases, drugs, chemicals, proteins, phenotypes, publications, and clinical trials. (d) A step-by-step example showing how the agent plans, queries knowledge bases, executes breadth-first research, collects observations, and uses programming actions to complete a systematic review task. (e) Benchmark results showing that DeepEvidence consistently outperforms strong baseline systems across four biomedical reasoning tasks.
• Figure 1: Gene-level correctness of four methods across evaluated targets in the target identification tasks. The heatmap reports binary correctness (colored as 0% or 100%) for each method: DeepEvidence, Biomni, ToolUniverse, and LLMs, on individual gene targets. Rows correspond to gene targets (with the number of test instances per gene shown in parentheses), and columns correspond to methods; cell colors indicate whether the method produced a correct prediction for that gene (green) or not (red). The rightmost column shows the average correctness across methods for each gene. Overall, DeepEvidence achieves the broadest and most consistent gene coverage, Biomni shows moderate performance, and ToolUniverse and LLMs succeed on fewer targets. Results are descriptive, as each gene–method pair is evaluated once.
• Figure 2: Experiment results of target identification. (a) Illustration of a target identification task that defines the research question, disease scope, target purposes, and methodological categories for therapeutic and diagnostic discovery. (b) Accuracy comparison for the target identification task showing that DeepEvidence achieves higher performance than baseline systems across evaluation settings. (c) An example end-to-end target identification workflow, where the agent conducts candidate search, broad evidence gathering, gene identity grounding across databases, evidence review and update, and comparative evaluation, progressively constructing and refining an evidence graph to identify and validate a disease-associated gene target.
• Figure 2: Pairwise recall comparison between agent-based systems and baseline LLMs in the evidence gap discovery tasks. Each panel plots task-level recall achieved by baseline LLMs (x-axis) against an agent-based system (y-axis): DeepEvidence (left), Biomni (middle), and ToolUniverse (right). Each point corresponds to one evaluation sample. The dashed diagonal indicates parity (y = x); points above the line denote cases where the agent outperforms the LLM, while points below indicate the opposite. Shaded regions highlight agent wins (green) and LLM wins (red). Insets report the win rate and counts across 20 reviews. DeepEvidence shows a substantial advantage over LLMs, whereas Biomni and ToolUniverse exhibit progressively weaker gains, highlighting differences in recall robustness across agent frameworks.
• Figure 3: Experiment results of mechanistic reasoning tasks. (a) Illustration of a mechanism of action and pathway reasoning task that specifies biological context, molecular entities, and reasoning objectives to explain signaling crosstalk and identify key regulatory checkpoints. (b) Accuracy comparison for the mechanism of action and pathway reasoning task showing that DeepEvidence achieves the highest performance compared with baseline systems. (c) Illustration of an in vivo metabolic flux response task where the agent reasons over tumor metabolic dependencies, target enzymes, and pathway vulnerabilities to identify cohorts with effective on-target flux suppression. (d) Accuracy comparison for the in vivo flux response task showing that DeepEvidence consistently outperforms baseline methods across evaluation settings.