Retrosynthesis Scoring with ChemProp Models

ISEF Category: Chemistry

Subcategory: Organic Chemistry  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

A molecule can look simple on paper and still be a nightmare to make. That is why chemists use retrosynthesis, a backwards planning trick that starts with the target and works back to safe, simple starting materials. If you can teach a model to score which routes look feasible, you can turn a chemistry puzzle into a machine learning project. That gives you a real way to compare green, kitchen-style routes against harder ones.

What Is It?

Retrosynthesis means planning a synthesis in reverse. You start with the product you want, then ask, “What smaller pieces could lead to this?” Think of it like tracing a finished necklace back to the beads and wire you need to build it.

ChemProp is a graph neural network for molecules. A graph neural network reads atoms as points and bonds as connections, then learns patterns from lots of examples. In this project, you would train a small model on reaction data from green chemistry subsets in the Open Reaction Database, then ask it to score how plausible different student-made routes look. The model does not replace a chemist, but it can rank ideas and reveal what features often show up in easier, cleaner reactions.

Why This Is a Good Topic

This topic works well for science fair research because it gives you a clear question, real data, and measurable output. You can test whether the model separates feasible routes from weak ones, and you can compare how performance changes when you use green chemistry examples, different molecular features, or different route labels. The project also connects to real problems in synthesis planning, safer chemistry, and AI for chemistry. You can learn data cleaning, model evaluation, and how to think like both a chemist and a data scientist.

Research Questions

• How does adding green chemistry reaction examples change model accuracy for retrosynthetic feasibility scoring?
• What is the effect of molecular size on predicted route feasibility scores?
• Does including atom economy related features improve route ranking compared with baseline molecular fingerprints?
• To what extent can a small ChemProp model distinguish plausible from implausible kitchen-synthesizable routes?
• Which reaction classes are easiest or hardest for the model to score correctly?
• How does class imbalance in the training set affect false positives for feasible routes?

Basic Materials

• Laptop with at least 16 GB RAM.
• Python installed with a scientific environment.
• Access to Google Colab or a local GPU if available.
• Open Reaction Database reaction files.
• Spreadsheet software for tracking samples and labels.
• Text editor or notebook for code notes.
• GitHub account for version control and file backup.

Advanced Materials

• University workstation or cloud GPU access.
• Python environment with PyTorch and ChemProp.
• Curated Open Reaction Database green chemistry subset.
• RDKit for molecular featurization and structure checks.
• SQL or data pipeline tools for reaction filtering.
• JupyterLab for model runs and analysis notebooks.
• Package for statistical testing, such as SciPy or statsmodels.

Software & Tools

• Python: Runs data cleaning, model training, and evaluation scripts.
• ChemProp: Trains molecular graph models for reaction or molecule prediction tasks.
• RDKit: Parses chemical structures, checks molecules, and generates descriptors.
• JupyterLab: Organizes code, notes, plots, and model comparisons in one place.
• Google Colab: Gives you free cloud computing when your laptop is too slow.

Experiment Steps

1. Define one clear feasibility label, then decide what counts as a plausible route and what counts as an implausible one.
2. Build a cleaned reaction dataset from Open Reaction Database entries, then filter it to a consistent green chemistry subset.
3. Choose the molecular representation you will compare, such as graph inputs, fingerprints, or simple descriptors.
4. Split the data so similar reactions do not leak across training and test sets.
5. Train a baseline model first, then compare it with your ChemProp model using the same evaluation metric.
6. Plan an error analysis that checks which molecules, reaction classes, or route types the model misses most often.

Common Pitfalls

• Mixing reaction types with different labels, which makes the model learn noise instead of feasibility patterns.
• Letting nearly identical reactions appear in both training and test sets, which inflates accuracy.
• Using raw database entries without cleaning salts, missing atoms, or inconsistent product names.
• Treating the model score as proof of real synthetic success, which overstates what the prediction means.
• Skipping class balance checks, which can make the model look good while missing rare but useful routes.

What Makes This Competitive

A strong version of this project goes beyond “does the model work.” You compare several input choices, test strict train-test splits, and explain why the model succeeds or fails on specific chemistry. You can also ask a deeper question, such as whether green chemistry examples improve route ranking for certain reaction families more than others. Careful error analysis and honest limits will make the project feel real, not just scripted.

Project Variations

• Use only oxidation and reduction reactions to see whether the model handles electron-transfer chemistry better than other classes.
• Compare green chemistry subsets with a broader reaction set to test whether cleaner reactions are easier to score.
• Test whether the model performs better on small, kitchen-synthesizable targets than on larger pharmaceutical-like molecules.

Learn More

• Open Reaction Database: Search the database for reaction records, filters, and metadata on reaction conditions and outcomes.
• MIT OpenCourseWare Organic Chemistry: Use course materials to review synthesis logic and reaction planning.
• PubChem: Look up structures, properties, and simple molecule data for candidate targets and intermediates.
• RDKit Documentation: Find free tutorials for molecule parsing, descriptors, and basic cheminformatics workflows.
• Nature Chemistry and Journal of Chemical Information and Modeling: Search for review articles and recent papers on retrosynthesis and molecular machine learning in PubMed or journal archives.