Predict Agrochemical Solubility With ChemProp Models

ISEF Category: Chemistry

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

The Hook

A molecule can look simple on paper and still refuse to dissolve in water. That matters because solubility affects how a chemical moves, works, and breaks down in the environment. You can build a model that predicts this behavior from structure alone, then use uncertainty to pick the next best compounds to test.

What Is It?

This project uses machine learning to predict aqueous solubility, which means how much of a compound can dissolve in water. ChemProp is a graph neural network model. Instead of treating a molecule like a string of letters, it reads atoms and bonds as a network, then learns patterns linked to solubility.

You start with SMILES, a text code that describes a molecule. Think of SMILES like a shorthand recipe for the structure. The model learns from known compounds, then predicts solubility for new ones. The uncertainty-aware part adds another layer. It tells you which predictions look shaky, so you can choose those compounds for follow-up validation first. That makes the project more efficient than testing random molecules.

Why This Is a Good Topic

This is a strong science fair topic because it blends prediction, data selection, and real chemistry. You can measure model quality with clear metrics, compare transfer learning against a baseline, and test whether uncertainty helps choose better follow-up compounds. The topic connects to pesticide behavior, environmental exposure, and how chemicals move through water. You can learn modern computational chemistry methods without needing a full synthesis lab for the first stage.

Research Questions

• How does transfer learning affect ChemProp performance on agrochemical solubility compared with training from scratch?
• What is the effect of training set size on solubility prediction error for agrochemicals?
• Does uncertainty-aware candidate selection improve the rate of successful wet validation hits compared with random selection?
• To what extent do molecular features such as polarity, halogen count, or ring system improve prediction accuracy?
• Which split strategy, random split or scaffold split, gives a more realistic estimate of out-of-sample solubility performance?
• How does adding related public solubility data change model calibration for uncommon agrochemical structures?

Basic Materials

• Computer with internet access and enough memory to run Python notebooks.
• Python installed with common data science libraries.
• Curated dataset of compounds with SMILES and measured aqueous solubility.
• Spreadsheet software for tracking compounds, predictions, and validation status.
• Access to a mentor or supervisor who can help coordinate wet-lab follow-up.

Advanced Materials

• Access to a university computing cluster or GPU workstation.
• Python environment with ChemProp, PyTorch, scikit-learn, and RDKit.
• Larger curated dataset of agrochemicals, plus external solubility datasets for transfer learning.
• Verified wet-lab solubility assay setup for candidate validation.
• ELN or database for versioned sample tracking and result logging.

Software & Tools

• ChemProp: Trains graph neural network models on molecular structures for solubility prediction.
• RDKit: Converts SMILES into molecular descriptors, fingerprints, and clean molecule records.
• Python: Runs model training, data cleaning, and evaluation scripts.
• scikit-learn: Calculates error metrics, cross-validation, and baseline models.
• Jupyter Notebook: Helps you explore data, document experiments, and compare model runs.

Experiment Steps

1. Define the target property, the compound class, and the exact solubility endpoint you will predict.
2. Gather a labeled dataset, then clean the SMILES strings, duplicate entries, and inconsistent measurement units.
3. Split the data in a way that tests real generalization, not just memorization of similar molecules.
4. Train a baseline model first, then compare it with a transfer-learned ChemProp model.
5. Build an uncertainty score and use it to rank which compounds deserve wet validation next.
6. Plan evaluation metrics that compare prediction accuracy, calibration, and the value of active selection.

Common Pitfalls

• Mixing solubility measurements from different assay conditions, which makes the labels noisy and hard to compare.
• Using a random split when closely related molecules appear in both train and test sets, which inflates performance.
• Keeping duplicate or near-duplicate SMILES in the dataset, which lets the model memorize rather than learn.
• Ignoring uncertainty calibration, which makes the model look confident even when it is wrong.
• Picking validation compounds only because they are easy to obtain, which weakens the active-learning test.

What Makes This Competitive

A competitive version of this project would do more than report one accuracy score. You would compare random selection against uncertainty-guided selection and show whether the model actually saves wet-lab effort. You would also test a harder split, like scaffold split, so the results reflect new chemistry instead of close look-alikes. Strong analysis of calibration, error patterns, and model limits would make the work feel much deeper.

Project Variations

• Use pharmaceutical solubility data instead of agrochemicals, then compare whether transfer learning helps one chemical class more than the other.
• Swap aqueous solubility for octanol-water partition behavior, then test whether the same model features still matter.
• Compare ChemProp with a simpler fingerprint model, then ask whether graph learning adds value for rare agrochemical structures.

Learn More

• ChemProp GitHub repository: Read the project documentation and example workflows on GitHub to understand model setup and training.
• RDKit Documentation: Learn how to parse SMILES and generate molecular descriptors, then find it through the RDKit project site.
• PubChem: Search compound records, molecular properties, and linked substance data for public chemical information.
• NIH PubMed: Search review articles on aqueous solubility prediction, molecular machine learning, and active learning in chemistry.
• MIT OpenCourseWare: Look for free courses in machine learning, data analysis, or computational chemistry to build background.