Drug Repurposing for Rare Pediatric Diseases

ISEF Category: Biomedical Engineering

Subcategory: Other  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

A drug that failed for one disease can sometimes help another. That idea has saved years of development time in medicine. For rare pediatric diseases, that shortcut can matter even more because patients often wait too long for new treatments. You can test whether a graph neural network can find those hidden matches better than simple ranking methods.

What Is It?

Drug repurposing means finding a new use for an old drug. Instead of starting from scratch, you look for links between drugs, genes, side effects, and diseases. A graph neural network is a machine learning model that learns from a network, not just a table. In this project, the network can connect drugs from DrugBank, drug-gene links from DGIdb, side effects from SIDER, and disease labels from Orphanet.

Think of it like a subway map. Each station is a drug, gene, side effect, or disease. Each track says two things are related. Your model looks for routes that lead from approved drugs to rare pediatric diseases. Then you compare its top predictions with later or ongoing ClinicalTrials.gov records to see whether the model pointed in the right direction.

Why This Is a Good Topic

This topic works well because it has a clear input, a clear output, and real data you can access without a wet lab. You can test whether a network model ranks known or emerging drug-disease links above chance, then compare it with simpler methods like shared target counts or similarity scores. The project connects to a real problem, rare pediatric diseases often have few treatment options and limited funding. You can also learn data cleaning, graph modeling, validation design, and how to judge a prediction against messy real-world outcomes.

Research Questions

• How does a graph neural network compare with shared-target scoring for ranking drug repurposing candidates for rare pediatric diseases?
• What is the effect of adding side effect links from SIDER on the model's ability to recover known drug-disease relationships?
• Does limiting the graph to pediatric disease nodes change prediction quality compared with a full rare-disease graph?
• To what extent do known ClinicalTrials.gov outcomes match the model's top-ranked repurposing candidates?
• Which feature set, drug-gene links, side effect links, or disease similarity, most improves prediction performance?
• How does training on older drug-disease links affect the model's ability to predict newer trial-tested candidates?

Basic Materials

• Laptop or desktop computer with at least 16 GB RAM.
• Python installed with Jupyter Notebook.
• Public database downloads from DrugBank, DGIdb, SIDER, Orphanet, and ClinicalTrials.gov.
• Spreadsheet software for cleaning and sorting metadata.
• Version control system such as Git for tracking code changes.
• Annotated list of known drug-disease pairs for validation.
• PDFs or exported tables from review articles on drug repurposing methods.

Advanced Materials

• Workstation or university cluster access with GPU support.
• Python with PyTorch Geometric or DGL.
• Network analysis tools such as NetworkX.
• SQL database or graph database for managing large merged datasets.
• PubMed abstracts for building a literature-based validation set.
• ClinicalTrials.gov API access or bulk downloads.
• Optional cheminformatics tools such as RDKit for drug similarity features.

Software & Tools

• Python: Cleans the source tables, builds the graph, and runs the model.
• Jupyter Notebook: Lets you test data prep, training, and evaluation step by step.
• NetworkX: Helps you inspect the structure of the drug-disease network before training.
• PyTorch Geometric: Supports graph neural network models for link prediction.
• ImageJ: Not needed for this topic, so skip it and focus on code-based analysis instead.

Experiment Steps

1. Define the prediction target, such as drug-disease links for rare pediatric conditions versus all rare diseases.
2. Merge the data sources into one graph and decide which node and edge types you will keep.
3. Choose a baseline method so you can compare the graph neural network against a simpler ranking approach.
4. Split known links into train, validation, and test sets in a way that prevents information leakage.
5. Design your evaluation plan around ranking metrics, calibration, and comparison with trial records.
6. Plan a validation table that separates confirmed matches, uncertain cases, and false positives.

Common Pitfalls

• Mixing adult and pediatric disease labels without a clear rule, which makes the prediction target too vague.
• Letting the same drug-disease pair appear in both training and test sets, which leaks the answer into the model.
• Using raw database names without standardizing identifiers, which creates duplicate nodes for the same drug or gene.
• Judging success only by top-ranked predictions, which hides whether the model performs better than a simple baseline.
• Treating later ClinicalTrials.gov entries as proof of efficacy, when many trials report only early-stage or incomplete outcomes.

What Makes This Competitive

A strong version of this project does more than run a model once. You need a careful baseline, clean data mapping, and a validation plan that tests whether the model truly beats simpler methods. The best entries often add a hard comparison, such as pediatric-only versus all-rare-disease training, or side-effect links versus no side-effect links. Strong analysis matters as much as model choice.

Project Variations

• Focus on only one rare disease group, such as neuromuscular or metabolic pediatric disorders, to see whether performance changes by subtype.
• Swap the graph neural network for a random forest or matrix factorization baseline and compare ranking quality.
• Add literature-mined gene-disease links from PubMed abstracts to test whether text evidence improves prediction.

Learn More

• PubMed: Search for review articles on drug repurposing, graph neural networks, and rare disease prediction.
• NIH Rare Diseases Clinical Research Network: Read about rare disease research needs and study design, then use it to frame your problem.
• ClinicalTrials.gov: Search for trial records tied to your predicted drugs and diseases, then compare predicted and observed outcomes.
• Orphanet: Use the disease registry and classification pages to define your rare pediatric disease set.
• MIT OpenCourseWare: Find free course materials on machine learning, graph algorithms, and computational biology.
• USGS ScienceBase or NOAA data portals: Not directly for this topic, but useful examples of how public data repositories organize large datasets.