Advancing molecular graphs with descriptors for the prediction of chemical reaction yields

Chemical yield can be increased using predictive algorithms to select high-yielding reactions. A recent study proposed a graph neural network architecture that combines structural information with molecular and reaction-level descriptors to predict yield.

The network generates reactants-product atom mapping and can even work with incomplete chemical reactions. The study compared different machine learning models and molecular representations and evaluated them based on classification and regression objectives. The dataset consisted of 10 reaction types, and the study was supplemented with an analysis of data, results, and errors.

Takeaways:

1. Chemical Yield Prediction: The study introduces a novel graph neural network (GNN) architecture for predicting chemical yield, improving reaction efficiency.
2. Structural & Molecular Features: The model integrates structural information, molecular descriptors, and reaction-level features, even handling incomplete reactions.
3. Comparison with ML Models: The GNN outperforms traditional models like logistic regression, SVM, CatBoost, and BERT-based approaches.
4. Molecular Representations: Various molecular encoding techniques, including fingerprints and SMILESVec embeddings, were assessed for yield prediction.
5. Dataset & Benchmarking: The models were trained on a proprietary dataset of 10 reaction types and tested on public datasets.
6. Advanced Analysis: The study explores steric effects, side reactions, and purification efficiency, with supplementary code available on GitHub.