BioKG Explorer: Interactive Visual Analytics for Biomedical Knowledge Graph Exploration over PrimeKG

Technical Summary: BioKG Explorer

Problem Statement
Biomedical knowledge graphs (KGs) like PrimeKG integrate vast, heterogeneous entities (diseases, drugs, genes, phenotypes, etc.) to support translational hypothesis generation. However, the scale and complexity of these graphs (PrimeKG contains ~129k nodes and ~4M relationships) create significant barriers to exploration. Existing tools often require query language expertise (e.g., Cypher), rely on static visualizations, or lack interactive, no-code interfaces capable of handling PrimeKG's specific scale. Furthermore, while Large Language Models (LLMs) are increasingly used for path explanation, current implementations typically treat the graph as a silent retrieval substrate or offer only chat-based interfaces, failing to integrate LLM-generated narratives directly with an interactive, visual graph canvas. There is a critical gap in open-source tools that combine full-scale PrimeKG exploration, bounded incremental subgraph visualization, and guarded LLM path explanations within a single, browser-based environment.

Methodology
BioKG Explorer is a full-stack web application designed to bridge these gaps through a three-layer architecture:

1. Data Layer (Neo4j): The system ingests PrimeKG via a reproducible notebook-based workflow. This process normalizes the raw CSV data into a property graph, preserving biomedical metadata and text descriptions. Relationships are deduplicated into undirected pairs, and nodes are enriched with searchable text fields and specific Neo4j labels (e.g., Disease, Drug, GeneProtein). Pre-computed analytics, including node degrees and disease similarity scores (based on shared genes, pathways, and phenotypes), are stored to facilitate rapid interaction.
2. Service Layer (FastAPI): A Python-based backend exposes RESTful endpoints for graph search, one-hop neighborhood expansion, top-k shortest path computation (using Neo4j Graph Data Science's Yen's algorithm), and disease-specific analytics. It manages the connection to Neo4j and returns JSON payloads containing graph structures and metadata.
3. Presentation Layer (React/D3.js): The frontend renders an interactive SVG graph canvas using D3.js. It supports incremental exploration where users can pin entities, expand one-hop neighborhoods, and manipulate the view without reloading the entire graph. Visual encodings map node types to colors and pre-computed degrees to node sizes.
4. Guarded LLM Integration: A unique feature is the optional path explanation module. When a user selects a shortest path, the backend constructs a bounded JSON context of the displayed subgraph. This context is sent to a configured Gemini model with strict system prompts instructing the LLM to summarize only the provided graph facts, explicitly avoiding clinical causality, diagnosis, or treatment recommendations. The resulting Markdown explanation is cached and displayed alongside the visual path.

Key Contributions
The paper presents BioKG Explorer as a tool contribution with four primary technical achievements:

• Ingestion Workflow: A pipeline that transforms PrimeKG into a Neo4j-compatible format while preserving semantic labels, relation types, and searchable text.
• Specialized API: A REST API supporting complex graph operations (search, expansion, shortest paths) and disease-centric analytics (candidate drugs, similar diseases) without requiring user knowledge of Cypher.
• Interactive Visualization: A browser-based interface featuring stable incremental layouts, undo/redo capabilities, and disease-specific evidence panels (candidate drugs and similarity scores) that render bounded subgraphs rather than the full graph.
• Guarded Explanation Workflow: A system that integrates LLM-generated path narratives directly with the visual graph, ensuring the explanation is grounded strictly in the user-visible subgraph context.

Results
The system successfully loads and manages the full scale of PrimeKG (129,375 nodes and 4,050,249 deduplicated relationships). The tool enables users to:

• Perform full-text searches to locate biomedical entities.
• Incrementally expand local neighborhoods to inspect context without performance degradation.
• Compute and visualize top-k shortest paths between source and target entities.
• Access disease-specific panels that list approved, off-label, and contraindicated drugs, as well as similar diseases with evidence counts.
• Generate natural-language explanations for selected paths that summarize the graph evidence without making clinical claims.

Significance and Claims
The author positions BioKG Explorer as a practical implementation pattern for ontology-aware visual analytics. Its significance lies in filling two specific gaps in the open-source ecosystem:

1. Providing an open-source, no-code, browser-based tool capable of interactive exploration at the full 4-million-edge scale of PrimeKG, distinguishing it from existing PrimeKG-native systems like TxGNN Explorer which are limited to pre-computed model predictions rather than open graph traversal.
2. Integrating LLM-narrated path explanations directly into an interactive graph canvas, moving beyond chat-only interfaces or static figure generation.

The paper explicitly states that the tool is designed for exploratory navigation, interpretation, and hypothesis generation, not for making clinical claims. It emphasizes that the system supports "graph-grounded" analysis, allowing researchers to move fluidly between entity search, local inspection, and path analysis. The author acknowledges limitations, noting that the tool inherits the biases of PrimeKG, that shortest paths do not imply causality, and that the LLM explanations are interpretive aids rather than biomedical authorities. The work serves as a foundation for future research into ontology-aware visual analytics and explainable AI in biomedicine.