Hsp90 Inhibitor Design with AI Tools

ISEF Category: Translational Medical Science

Subcategory: Drug Identification and Testing  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

Cancer cells often lean hard on Hsp90, a protein that helps other proteins stay folded and working. If you can design a molecule that blocks the right part of Hsp90, you may shut down a disease pathway without hitting the wrong site. Your project can test whether AI-made molecules prefer the N-terminal pocket over the C-terminal pocket. That makes this a real drug-design problem, not just a coding exercise.

What Is It?

Hsp90 is a helper protein called a chaperone. Think of it like a backstage crew that keeps other proteins from falling apart. Many cancer cells depend on Hsp90 more than healthy cells do, so researchers look for inhibitors, which are molecules that block Hsp90’s job.

Your project focuses on the N-terminal domain, one binding site on Hsp90, while avoiding the C-terminal pocket, another site that can cause unwanted effects. You are not trying to make a drug in the lab. You are using computer models to propose molecules, score how well they might fit the target, and filter out ones that look weak or risky. That makes the project a mix of chemistry, biology, and machine learning.

Why This Is a Good Topic

This topic works well for a science fair because you can ask clear, measurable questions about selectivity, docking score, and predicted drug-likeness. You also connect to a real medical problem, which is finding safer cancer-targeted inhibitors. Even if you are new to research, you can learn how to search protein structures, compare candidate molecules, and judge models with data instead of guesses.

Research Questions

• How does conditioning a generative model on Hsp90 N-terminal binding change the docking score of proposed molecules?
• What is the effect of adding C-terminal off-target filters on the chemical diversity of generated inhibitors?
• Does REINVENT produce more selective Hsp90 candidates than MolGPT under the same input constraints?
• To what extent do top-scoring generated molecules pass ADMET-AI drug-likeness filters?
• Which molecular features best predict separation between N-terminal and C-terminal docking scores?
• How does the choice of training set affect the novelty of generated Hsp90 inhibitors?

Basic Materials

• Laptop or desktop computer with stable internet access.
• Google Colab account with free GPU access when available.
• PubChem compound records for known Hsp90 inhibitors.
• Protein Data Bank structures for Hsp90 N-terminal and C-terminal domains.
• Python notebook for data cleaning and analysis.
• Spreadsheet software for tracking candidate molecules and scores.

Advanced Materials

• University or cloud access for larger model runs.
• PyTorch environment for running REINVENT or MolGPT.
• RDKit for molecular cleaning, descriptor calculation, and filtering.
• DiffDock for protein-ligand docking predictions.
• ADMET-AI or a similar open prediction pipeline for absorption, distribution, metabolism, excretion, and toxicity estimates.
• Molecular visualization software such as PyMOL or UCSF ChimeraX.
• GPU-enabled workstation or cluster access for faster model training.

Software & Tools

• Google Colab: Runs notebook-based chemistry models without local setup.
• Python: Handles data cleaning, scoring, and visualization.
• RDKit: Filters molecules, computes descriptors, and checks chemical validity.
• DiffDock: Predicts how candidate molecules may bind to Hsp90 pockets.
• ImageJ: Not needed for the chemistry itself, but useful if you later compare assay images with follow-up wet-lab work.

Experiment Steps

1. Define the exact comparison you want, such as N-terminal selectivity, novelty, or predicted ADMET balance.
2. Assemble a clean reference set of known Hsp90 ligands and decoys so your model has a fair starting point.
3. Choose one generative pipeline and one scoring pipeline, then keep the input rules consistent across runs.
4. Set up a filtering plan that rejects invalid structures, weak predicted binders, and molecules with poor drug-likeness flags.
5. Build a ranking system that combines docking, selectivity, and ADMET outputs into one clear decision rule.
6. Plan a validation test with held-out molecules so you can show your model works on data it did not see before.

Common Pitfalls

• Training on mixed Hsp90 datasets without separating N-terminal and C-terminal binders, which blurs the selectivity signal.
• Treating a strong docking score as proof of real binding, which ignores model error and false positives.
• Forgetting to standardize molecule formats, which makes duplicate structures and salt forms look like different compounds.
• Overfiltering for ADMET early, which can remove novel candidates before you compare chemical space.
• Comparing model outputs from different runs without fixing seeds or selection rules, which makes your results hard to reproduce.

What Makes This Competitive

A stronger version of this project goes past simple docking lists. You would compare models with held-out validation, report selectivity between two Hsp90 pockets, and show how each filter changes the chemistry that survives. You could also test whether your ranking method keeps novelty while improving predicted safety. That mix of model design, controls, and quantitative comparison can make the project feel like real discovery work.

Project Variations

• Use a different target class, such as another chaperone protein, to see whether the same generative pipeline still finds selective binders.
• Compare docking-only ranking with docking plus ADMET filtering to measure how much each step changes the final candidate set.
• Swap the input library from known inhibitors to fragment-like molecules and test whether the model still builds valid Hsp90 candidates.

Learn More

• PubChem: Search compound records for known Hsp90 inhibitors, structures, and assay summaries.
• RCSB Protein Data Bank: Find 3D structures of Hsp90 domains and bound ligands.
• National Center for Biotechnology Information Bookshelf: Read free review chapters on molecular chaperones and drug discovery.
• PubMed: Search for review articles on Hsp90 inhibition, docking, and structure-based drug design.
• MIT OpenCourseWare: Look for free materials on biological chemistry, molecular biology, or drug discovery methods.
• NIH: Search for free resources on pharmacology, toxicology, and translational research topics.