ADInt Explorer

An interactive visualization dashboard for the ADInt knowledge graph : the Alzheimer's drug-repurposing knowledge graph from Xiao et al., Scientific Reports 2024. Built as the final project for HINF 5620 at the University of Minnesota.

Recommended way to explore this project: click the Open in Colab badge above. The notebook walks through the full pipeline from raw data files to live interactive dashboard, with structured headings and a clickable Table of Contents.

What it does

The Zhang lab's ADInt knowledge graph contains over a million literature-extracted relationships between drugs, dietary supplements, complementary health interventions, diseases, and genes — with Alzheimer's Disease at its center. The published artifact is a static tab-separated file. ADInt Explorer is the interactive interface that turns it into a navigable research instrument.

Four core features:

• Search & refocus : type any of the 162,000 concepts; the graph re-centers around it instantly.
• Category filters : toggle the seven Zhang lab categories on or off; rare types like CIH always survive thanks to per-category stratification.
• Mechanistic path finder : pick any two concepts; NetworkX computes the shortest chain through 742,000 edges in under a second and highlights it in gold.
• Evidence panel : click any edge to surface the exact source sentence behind every triple, with a clickable PubMed link. Every connection traces back to literature.

How to run

Recommended: Google Colab (no installation)

Click the Open in Colab badge at the top of this README. The notebook contains the full pipeline:

1. Section 1 installs dependencies and mounts your Google Drive
2. Section 2 lets you point at your data folder
3. Sections 3–10 walk through the data pipeline, exploratory visualizations, and app construction
4. Section 11 launches the live dashboard and prints a clickable URL

Total time after the first cell: ~1 minute including data load.

Alternative: Run locally

git clone https://github.com/YOUR-USERNAME/adint-explorer.git
cd adint-explorer
pip install -r requirements.txt

# Download data from the Zhang lab repository:
# https://github.com/zhang-informatics/ADInt
# Place ADIntKG.tsv and Neo4j_node_updated.csv in the project root,
# preserving the Neo4j_data/ subfolder.

python app.py
# Open http://127.0.0.1:8050 in your browser

Repository structure

adint-explorer/
├── app.py                         Standalone Dash application
├── generate_slides.py             Builds the project presentation deck
├── requirements.txt               Pinned dependencies
├── notebooks/
│   └── ADInt_Explorer.ipynb       RECOMMENDED: full walkthrough notebook
├── figures/
│   ├── screenshot_dashboard.png
│   ├── fig_architecture_flowchart.png
│   ├── fig_degree_by_category_with_outliers.png
│   └── fig_ad_neighborhood_structure.png
├── README.md
└── LICENSE

Architecture

The system has two phases. Build phase runs once at startup: load the source files, deduplicate triples, build a NetworkX graph, and pre-index evidence by (subject, predicate, object) for O(1) edge-click lookups. Interaction loop runs on every user action: search, click, path-find, or filter triggers a recomputation of a small focal subgraph (top-7 neighbors per category) which is then sent to the browser for rendering.

Design rationale

The visualization design is grounded in Dashboard Vision (Yang et al., IEEE TVCG 2025). Four guidelines applied:

• L1 — stratified layout → category-wedge wheel guarantees rare types appear in every view
• O2 — big numbers as primary focal points → live counts in the status bar
• O1 + O4 — subtitles and inline labels drive deepest attention → source sentence placed directly beside the edge it explains
• L3 — simplified groupings → focal-only edge filter reduces clutter ~63% by default

Knowledge graph structure

Two visualizations from the notebook (Section 6) explain why the design works the way it does:

Long-tailed degree distribution per category — Alzheimer's Disease at degree 8,394 dominates	AD's edges by predicate × neighbor category — PREVENTS is sparse (~4%); CIH is barely visible

Data

This tool requires the published ADInt knowledge graph data, available at github.com/zhang-informatics/ADInt. The data files are not included in this repository.

Citation

If you use this tool in research, please cite the underlying ADInt paper:

Xiao, Y., Hou, Y., Zhou, H., Diallo, G., Fiszman, M., Wolfson, J., Zhou, L., Kilicoglu, H., Chen, Y., Su, C., Xu, H., Mantyh, W. G., & Zhang, R. (2024). Repurposing non-pharmacological interventions for Alzheimer's disease through link prediction on biomedical literature. Scientific Reports, 14, 8693. doi:10.1038/s41598-024-58604-8

Future directions

• Predicate-based filtering — let users hide correlative predicates (ASSOCIATED_WITH, COEXISTS_WITH) to focus on therapy and causation claims
• DrKGC integration — overlay the Zhang lab's R-GCN link-prediction confidence scores as edge weights
• Evidence depth filter — show only edges supported by ≥ N PubMed papers
• ClinicalTrials.gov annotation — flag intervention nodes in active Alzheimer's trials
• Public deployment — host on Render or Hugging Face Spaces for browser-only access

Author

Aviral Bhatnagar · Health Informatics PhD · University of Minnesota · bhatn042@umn.edu

License

MIT — see LICENSE.