Neurodegenerative Disease Knowledge Graph Explorer

An NLP tool that pulls real research abstracts from PubMed, extracts typed subject-verb-object relationships using spaCy's dependency parser, cross-references active ClinicalTrials.gov studies, and visualizes it all as an interactive, color-coded knowledge graph — built as a self-directed exploration of applying NLP to neurodegenerative disease research.

What it does

1. Pick a disease preset (Alzheimer's, Parkinson's, ALS, Huntington's, FTD, Lewy body dementia) or enter a custom PubMed search term
2. Pulls matching abstracts live from PubMed via NCBI's Entrez API, plus a plain-language Wikipedia overview and active ClinicalTrials.gov studies
3. Extracts relationship triples like (amyloid-beta, disrupt, neuronal function), each typed by entity category (disease / gene-protein / drug-chemical / anatomy / process), flagged for negation ("X does not activate Y"), and linked back to its source PMID
4. Renders the relationships as an interactive, color-coded knowledge graph, with a searchable table, CSV/GraphML export, and the underlying source abstracts alongside the recruiting trials

Try it

Live app: (add your Streamlit Community Cloud link here once deployed)

Run it yourself

pip install -r requirements.txt
streamlit run app.py

How it works

• Data sources: NCBI's free PubMed E-utilities API (via Biopython's Entrez + Medline modules), ClinicalTrials.gov's public API v2, and Wikipedia's REST summary API
• NLP: spaCy's en_core_web_sm dependency parser for triple extraction (nlp_utils.py), with a curated neurodegenerative-disease lexicon for entity typing/synonym normalization and a negation detector
• Graph: NetworkX for graph structure, pyvis for the interactive rendering
• Interface: Streamlit, deployed on Streamlit Community Cloud

Known limitations

The subject-verb-object extraction is a beginner-level heuristic based on dependency parsing, not a trained relation-extraction model — it misses complex or passive-voice sentences and occasionally produces a noisy triple. Entity typing is lexicon-based rather than a trained biomedical NER model (e.g. scispaCy), so anything outside the curated term list falls back to a generic "OTHER" category. Improving either (a proper relation-extraction model, or swapping in scispaCy for entity typing) would be a natural next step.

Disclaimer

This is a research-exploration and portfolio project, not a medical or diagnostic tool. Extracted relationships are computed automatically and may contain NLP errors — always verify against the original source abstract.

Motivation

Built while exploring the intersection of NLP, knowledge graphs, and neurodegenerative disease research.