Homogeneous catalyst graph neural network: A human-interpretable graph neural network tool for ligand optimization in...

Area of Science:

• Catalysis
• Machine Learning
• Computational Chemistry

Background:

• Traditional metal-ligand catalyst optimization relies on inefficient empirical trials.
• Developing novel, highly selective asymmetric catalysts is crucial for chemical synthesis.

Purpose of the Study:

• To introduce HCat-GNet, a machine learning model for predicting enantioselectivity in asymmetric catalysis.
• To provide an interpretable method for identifying key ligand structural features influencing catalyst performance.
• To demonstrate the model's ability to extrapolate to novel ligand structures and generalize across different reactions.

Main Methods:

• Development of the Homogeneous Catalyst Graph Neural Network (HCat-GNet) model.
• Training the model using SMILES representations of molecules to predict reaction enantioselectivity.
• Utilizing interpretability features to identify influential atoms within ligands.
• Validation on a new class of rhodium-catalyzed asymmetric 1,4-addition ligands and benchmark datasets.

Main Results:

• HCat-GNet accurately predicts enantioselectivity for asymmetric reactions.
• The model provides atom-level insights into ligand contributions to selectivity.
• Successful extrapolation to novel chiral ligand space was demonstrated.
• Generalizability across different asymmetric reactions was confirmed.

Conclusions:

• HCat-GNet offers an efficient, data-driven approach to ligand optimization in asymmetric catalysis.
• The model's interpretability facilitates rational catalyst design.
• HCat-GNet represents a significant advancement in computational catalyst development.