Recent Applications of Machine Learning in Molecular Property and Chemical Reaction Outcome Predictions

Area of Science:

• Computational Chemistry
• Cheminformatics
• Artificial Intelligence

Background:

• Machine learning (ML) is increasingly adopted in chemistry, building on successes in molecular property prediction (MPP) and chemical reaction prediction (CRP).
• Existing ML applications range from ensemble methods like random forests to sophisticated graph neural network algorithms.
• Feature engineering and learning approaches are crucial for enhancing ML model performance in chemical tasks.

Purpose of the Study:

• To provide a non-mathematical overview of recent ML applications in predictive chemistry.
• To highlight advancements in ML for both molecular property prediction (MPP) and chemical reaction prediction (CRP).
• To discuss challenges and future prospects of ML in chemical research.

Main Methods:

• Review of ML implementations, including ensemble-based random forest models and graph neural networks.
• Exploration of feature engineering and feature learning techniques used with ML models.
• Analysis of reported accuracy metrics (e.g., RMSE) for MPP and CRP tasks.

Main Results:

• ML models like D-MPNN, MolCLR, SMILES-BERT, and MolBERT achieve high accuracy in MPP tasks (e.g., predicting lipophilicity, solubility).
• ML applications in CRP face challenges due to the complexity of multiple interacting molecules.
• Reported RMSEs for MPP range from 0.287 to 2.20, while CRP yield predictions often exceed 4.9, reaching over 10.0 in some cases.

Conclusions:

• ML offers powerful tools for accelerating molecular design and drug discovery through accurate MPP.
• Significant challenges remain in handling complex reaction datasets for CRP.
• ML-driven workflows present an optimistic future for advancing predictive chemistry in both MPP and CRP domains.