Reaction Center Prediction by Analyzing Attention of a Chemical Language Model

Key Points

• The aim is to understand how attention mechanisms in pretrained chemical language models contribute to predicting molecular properties and reactions.
• Analyzed attention matrices of pretrained chemical language models.
• Investigated correlations between attention values and prediction accuracy for molecular properties and reactions.
• Identified key attention patterns that correlate with accurate predictions of molecular properties.
• Demonstrated that certain attention mechanisms significantly enhance the capability of chemical language models in reaction prediction.