koredami/misato-electron-affinity

MISATO Electron Affinity Dataset (Curated Subset)

This dataset is a curated subset of the MISATO dataset, focused on predicting electron affinity using quantum molecular descriptors.

It includes:

• train.csv and test.csv with quantum features for ligands
• Model training and evaluation scripts (train_custom_model.py, hyperparameter_tuning.py)
• A visualization of predictions (prediction_vs_true.png)

Features

Each molecule is described by:

• Ionization Potential
• Hardness
• Electronegativity
• Koopman's Estimate
• Molecular Weight
• Total Charge

The target variable is:

• Electron Affinity

Use Case

The dataset is intended for training and evaluating regression models to predict electron affinity, a key property for drug discovery and chemical reactivity analysis.

Files

File	Description
train.csv	Training dataset (80%)
test.csv	Test dataset (20%)
train_custom_model.py	Baseline model training
hyperparameter_tuning.py	Grid search tuning
build_dataset.py	Script to curate the dataset from QM.hdf5
prediction_vs_true.png	Plot of predicted vs. true values

Results Summary

Metric	Value
Test RMSE	0.0464
Model	RandomForestRegressor (scikit-learn)

License

This dataset is shared under the CC BY 4.0 license.
Original data from MISATO on Zenodo.

Final Report

You can view the full technical report describing the dataset, model development, results, and evaluation here: View Report